Refactor `FederatedServer` and `FederatedClient` code.

This Merge Request implements a number of revisions to the declearn.main submodule and most importantly to the FederatedClient and FederatedServer classes, that aim at making the code easier to read, maintain, test and override, as well as making it more end-user-friendly.

Most of these changes were first implemented on a working branch as part of the effort towards implementing Differential Privacy in declearn (see MR !11 (merged)). A combination of git cherry-picking and of further implementation efforts were used to make up the current MR, that paves the way towards merging the already-developed local DP features within the revised code.

Summary of the modifications implemented here:

• Implement some automation tools to wrap up configurations:

  • Add a generic TomlConfig ABC, that enables writing heavy parameter containers and parsing TOML files.
  • Add some generic dataclass-generation tools.

• Revise FederatedServer:

  • Implement FLRunConfig based on TomlConfig, that wraps together server-side "run" method parameters.
  • Move the former as well as existing server-oriented dataclasses to the new declearn.main.config submodule.
  • Have initialization make use of the full config, while training and evaluation only access their own section.
  • Clean up some backend code regarding client results' collection, to improve readability and reduce redundancy.

• Revise FederatedClient:

  • Split out the training and evaluation routines into a new TrainingManager class (under declearn.main.utils).
  • Write some unit tests on the former class, that are designed to verify that effort constraints work properly.
  • Fix a bug where using epochs=None, steps=... training constraints would result in an infinite loop.

declearn/main/config/__init__.py

+# coding: utf-8
+
+"""Tools to specify hyper-parameters of a Federated Learning process.
+
+This submodule exposes dataclasses that group and document server-side
+hyper-parameters that specify a Federated Learning process, as well as
+a main class designed to act as a container and a parser for all these,
+that may be instantiated from python objects or from a TOML file.
+
+In other words, `FLRunConfig` in the key class implemented here, while
+the other exposed dataclasses are already articulated and used by it.
+"""
+
+from ._dataclasses import (
+    EvaluateConfig,
+    RegisterConfig,
+    TrainingConfig,
+)
+from ._run_config import FLRunConfig