Add 'TrainingManager.(train|evaluate)_under_constraints' to the API.

- Until now, the public methods for training and evaluation using
  the `declearn.main.utils.TrainingManager` class (and its DP-SGD
  counterpart) required message-wrapped inputs and emitted similar
  outputs.
- With this commit, the previously-private `train_under_constraints`
  and `evaluate_under_constraints` routines are made public and thus
  part of the declearn API, with some minor refactoring to make them
  more user-friendly.
- The rationale of this change is to enable using `TrainingManager`
  outside of our `FederatedClient`/`FederatedServer` orchestration,
  notably when simulating FL training or testing client-side code.
  It may also be helpful to end-users that would like to build on
  declearn but implement their own orchestration tools or algorithm
  loops.
- In the future (declearn >=3.0), we may go one step further and
  take all the messaging-related instructions out of the current
  class. The class may also be moved to a different namespace, e.g.
  a new 'declearn.train' module; but this is entirely out of scope
  for now.