It seems like GPU support is broken again due to conflicting version
installation. This does not mean that our code is broken, but for
now we will drop testing it as part of the CI/CD.

Finalize version 2.6.0

See merge request !70

- Add SecAgg, Fairness and evaluation-frequency options to the usage
  guide.
- Add SecAgg and Fairness related phases to the FL process overview.
- Add Fairness API to the package overview.

Enable Fairness-Aware Federated Learning

See merge request !69

- Fairness rounds occur prior to training ones. Consequently, this
  commit modifies their indexing to be aligned with that of the
  _previous_ training round. As a result, utility and fairness
  metrics are properly aligned in checkpoint files.
- In addition, this commit makes it so that a final fairness
  evaluation round is run prior to ending the FL process when
  checkpointing is setup at the server level.

This generic constructor enables parsing any subclass instance from
serializable specifications. This is notably useful to ensure that
`FLOptimConfig` remains compatible with TOML parsing even when a
fairness algorithm is used.

Dedicated tests and further documentation are yet to be added.

- Always have clients send back an 'InitReply' after the first
  'TrainingManager' setup step happened without error.
- Then, rely on step-specific query-reply exchanges for further
  steps (SecAgg, DP-SGD and/or Fairness setup).

When a '__all__' value is provided in a public module, make use of
it instead of collecting each and every member.

This fixes the collection breaking upon encountering external
packages such as 'warnings' in an '__init__.py' file.