Fix the case when the batch size for a given subset is
exactly a multiple of the number of samples for that
subset, which would result in an exception (at least
with in-memory data).

- Add shared code in the unit tests suite.
- Add dedicated data and tests in FairGrad and FairBatch subclasses.

- Expose some subroutines under setup and fairness round,
  for the mere sake of making tests easier to perform, as
  well as to enable variants over current algorithms in
  the future / in experiments.
- Rename some methods and re-order some arguments.
- Refactor server-side aggregation of metrics, making it
  part of 'FairnessControllerServer' rather than part of
  'FederatedServer' backend code.

- Replace 'FairfedFairnessFunction' with 'FairfedValueComputer'.
- Add explicit 'target' keyword argument to the controller classes.

- Revise 'FairnessAccuracyComputer' into 'FairnessMetricsComputer'.
  - Revise name and documentation to reflect broader scope.
  - Remove specific accuracy and loss computation routines.
  - Introduce more generic metrics-computation routines.
  - Add methods to set up accuracy and loss-computation metrics.
- Revise 'FairnessControllerClient.compute_fairness_measures':
  - Have unshared values be returned as a nested dict,
    to limit redundant (un)packing operations and code.
  - Add 'setup_fairness_metrics' method to define metrics
    to compute. By default, compute group-wise accuracy.
  - By default, scale and share all metrics defined by
    `setup_fairness_metrics`, and compute (but do not
    share) local fairness values.
- Propagate this change across all subclasses, effectively
  trimming some redundant boilerplate code.

- Until now, 'TomlConfig' fields were either mandatory of had a
  fixed default value. In some cases, the latter may be refined
  based on other values; however, they could not be missing in
  TOML files.
- With the added 'autofield_fields' class attribute, some fields
  can be explicitly marked as having a dynamically-created default
  value. In that case, it can safely be ignored in TOML files.
- This is now used to support omitting 'evaluate' and 'fairness'
  fields when writing down a 'FLRunConfig' as a TOML file.

Debug 'FederatedClient' along the way, mostly improving error
catching and raising behaviors' consistency with other methods.

- Move all ABCs and the ever-present accuracy computer to 'api'.
- Make 'core' a submodule for concrete implementations of abstract
  classes, currently holding group-fairness functions and in-memory
  fairness dataset.

- Introduce a 'strict' mode, that sticks closely to the original
  paper, restricting the scope of settings covered by FairFed
  and using rather-specific synthetic fairness values.
- Document this 'strict' mode and, by contrast, the more generic
  extension that was previously implemented.

This implementation is very loosely based on the "legacy" one from
the 'fairgrad' branch. It generalizes to broader cases than those
showcased in the initial FairFed paper, and relies on a dedicated
(mostly client-side) Aggregator subclass that is altered during
fairness rounds to use fairness-based averaging weights.

The maths to the "global" fairness measures computation are not
those of the original paper, and should be verified to equate the
original paper for the EOD and SPD cases  prior to merging.

- Have clients' measures-computing routine return both shareable
  and local-use values.
- This way, unscaled metrics that are to be post-processed and/or
  checkpointed can be kept as-is rather than scaled-then-descaled.

This commit adds the Fed-FairBatch and FedFB algorithms for group
fairness. It is based on the "legacy" implementation from the
'fairgrad' branch, but adjusts to the updated fairness API.

It also adds some documentation, changes the signature of sampling
probabilities controllers to be coherent with 'FairnessFunction',
and integrates FedFB as a variant of Fed-FairGrad, within the same
controller class.

Co-authored-by: Brahim Erraji <brahim.erraji@inria.fr>

- Revise the way client-side controllers are instantiated from
  server-emitted instructions, using the usual type-registration
  tools and removing the need to subclass setup query messages.
- Move fairness-related base messages to the 'messaging' module
  (and 'secagg.messaging' one).
- Have the client-side controller access and wrap a training
  manager at instantiation. Revise method signatures and docs.
- Add fairness metrics checkpointing.

* Integrate Fairness configuration on the server side, split
  between 'FLOptimConfig' and 'FLRunConfig'.
* Implement generic fairness measures computation on the server
  side. Implement default computations on the client side. Leave
  additional actions up to subclasses.
  This will potentially be refactored to be coherent as to what
  goes in the fairness controllers and what goes in the existing
  main classes, and as to what is part of the shared API and what
  is left to algorithm-specific subclasses.
* Move fairness round query and reply messages to 'messaging'.