- Run `Aggregator` unit tests for available frameworks only.
- Skip `gRPC`-dedicated tests when grpc is not installed.
- Skip `DPTrainingManager` tests when opacus is not installed.
- Fix toy regression functional tests failing when some frameworks
  are missing due to the use of too-specific type hints.

Recent JAX versions have deprecated, then dropped, the `device()`
method of jax (numpy) arrays, in favor of the already-existing,
distinct-return-type `devices()` one.

This commit therefore replaces calls to `device` with operations
that include calling `devices` and extracting its content. This
is not very elegant, but seems to be the proper way to do things
based on the current JAX documentation.

Finalize DecLearn v2.4.0.

See merge request !63

- Apply some minor backward changes (and, when required, branches) to
  support breaking changes in Keras 3 (the default for TensorFlow 2.16
  and onwards), while preserving support for older versions.
- On the side, split a test about 'Model' serialization in two tests,
  one for config serializability, the other for instantiation from a
  non-serialized config dict.
- The current code has been tested (by running all unit tests) with
  both TensorFlow 2.11 and 2.16, in a Python 3.10 environment.

- Add new 'declearn.communication.utils' submodule.
- Move existing utils there, but keep re-exporting them from the
  parent level to avoid breaking existing code.
- Refactor some backend code used to verify exchanged messages'
  type as part of federated learning processes into two new utils.

- Until now, in the main FL process, clients would always receive
  global model weights twice, due to their being shared both with
  training and evaluation requests.
- This commit fixes that via the introduction of backend mechanisms
  to keep track of which clients hold the latest model weights, and
  sparingly attach them to requests based on this.
- As a side effect, `TrainRequest` and `EvaluationRequest` messages
  now specify `weights` to be an `Optional[Vector]` field.

This commit introduces imports of 'tf.keras' as 'tf_keras' so that
false-positive pylint errors can be silenced in one place per file
rather than on each and every call.
This remains tiresome, but might in fact be useful to switch to the
newer TensorFlow versions in the future.

Set minimal version to 2.11.0 (rather than 2.5.0):
- 'TensorflowOptiModule' adapter is based on the "new" Optimizer API
  introduced with TensorFlow/Keras v2.11, but the required dependency
  version had not been updated until now.

Set maximal version to 2.15.X:
- TensorFlow 2.16 (the first release candidate of which was added to
  PyPI three days ago) is switching to Keras 3 by default, which is
  bound to introduce breaking changes (current unit tests fail with
  'tensorflow==2.16.0rc0').
- For now, the maximal version is thus set to 2.15.X. This may change
  in the near future when some time and effort is put in understanding
  changes and/or configuring dependencies to keep using Keras 2 rather
  than Keras 3 in TensorFlow's backend.

- The previous implementation would result in dropping some kwargs
  (namely, 'name' and 'queue') when running a function as process.
- The new backend implementation prevents this using a wrapped
  instead of a partial function (still relying on 'functools').

- Document changes in the registration and initialization processes.
- Document changes of the Optimizer AuxVar API.
- Document requirement for peers using the same DecLearn version.

Revise network communication backend and message-parsing logic.

See merge request !62

Set up a low heartbeat for localhost-running tests, resulting in
drastic runtime gains (notably due to time gains during the join
and initialization phase for functional tests).

- Update the way received messages are type-checked and have them
  be deserialized after type-checking only.
- Move up metadata collection and use to a new optional and sparse
  step during the initialization phase.
- Implement new message subtypes for metadata query and reply, and
  for signalling that initializations steps went well.

- Revert some changes from a previous commit, that may be envisioned
  again when revising APIs in-depth for DecLearn 3.0.
- Stitch back the 'communication' and 'messaging' submodules together,
  having 'NetworkClient' and 'NetworkServer' send 'Message' instances
  (that are merely serialized to string prior to being exchanged) and
  parse received strings into 'SerializedMessage' instances.
- This change is a compromise between keeping things as before (which
  would not benefit from the introduced delayed-parsing capability of
  'SerializedMessage'), and changing them so deeply that current code
  making use of communication endpoint would require heavy revisions.

* Introduce new 'declearn.messaging' submodule, that is meant to
  replace the now-deprecated 'declearn.communication.messaging'.
* Make 'Message' extendable via subclassing and type-registration
  (which is automatically performed by default).
* Introduce 'SerializedMessage' as a wrapper for received messages,
  that parses the exact message subtype (enabling logic tests and
  message filtering) but delays actual content deserialization and
  Message object recovery (enabling to prevent undue resources use
  for unwanted messages that end up being discarded).
* Deprecate some old 'Message' classes, that are still exported by
  the deprecated submodule but not by the new one.

- Introduce a new 'declearn.communication.api.backend' submodule:
  - Introduce a new 'ActionMessage' minimal backend API, that defines
    hard-coded, lightweight and easily-parsed data structures designed
    to convey information and content across network communications
    agnostic to the content's nature.
  - Move 'declearn.communication.messaging.flags' submodule to
    'declearn.communication.api.backend.flags'.
  - Move 'MessagesHandler' semi-private class to the 'backend'
    submodule, where it is now properly exposed.

- Revise 'MessagesHandler':
  - Have it use the new 'ActionMessage's to wrap up messages, and
    delay their content's parsing to other times and parts of the
    code.
  - Quit sending information together with registration requests.
    In the future, we may want to introduce some callback system to
    validate or complete a registration request's processing.
  - Turn 'heartbeat' into an instantiation parameter.

- Revise 'NetworkClient' and 'NetworkServer':
  - Quit passing of 'data_info' with registration requests.
  - Have message-sending and message-receiving methods expect str
    inputs and return str outputs: the actual parsing into objects
    is therefore delayed and left to be defined elsewhere.
  - Rename 'NetworkClient.check_message' into 'recv_message' (keep
    the former as an alias, with a DeprecationWarning).
  - Improve the use of (optional) timeouts when sending or expecting
    messages and overall exceptions handling:
    - 'NetworkClient.recv_message' may either raise a TimeoutError
      (in case of timeout) or RuntimeError (in case of rejection).
    - 'NeworkServer.send_messages' and 'broadcast_message' quietly
      stops waiting for clients to collect messages after the (opt.)
      timeout delay has passed. Messages may still be collected.
    - 'NetworkServer.wait_for_messages' no longer accepts a timeout.
    - 'NetworkServer.wait_for_messages_with_timeout' implements the
      possibility to setup a timeout. It returns both received client
      replies and a list of clients that failed to answer.