discojs documentation update

[leonard-amsler]

The documentation of discojs is no longer aligned with the codebase. Classes no longer exist (DistributedTrainer, LocalTrainer), methods are renamed (fit becomes train) and outdated aggregators. No code change in this PR.

Changes:

Trainer:

• main method is train not fit, and it trains through model.train.
• It is no longer abstract and does not contain onRoundBegin/onRoundEnd callbacks. It calls the client's onRoundBeginCommunication and onRoundEndCommunication for communication with the server/other clients.
• As it is not abstract anymore, the DistributedTrainer and LocalTrainer no longer exist.
• Documented stopTraining. It lets the current round finish and disable future ones.
• Added a section about differential privacy applied before weights are shared with link to PRIVACY.md.

Client:

• Three implementations instead of two: FederatedClient, DecentralizedClient and LocalClient.
• Clarification that the client does not aggregate anything.

Aggregators:

• Three strategies instead of two: MeanAggregator, SecureAggregator and ByzantineRobustAggregator (centered clipping with per-client momentum).
• Stated which ones are available per protocol: all three for decentralized, MeanAggregator and ByzantineRobustAggregator for federated, matching TrainingInformation.networkToSchema.

Disco:

• disco.fit(data) becomes disco.trainFully(data), and Trainer.fitModel() → Trainer.train() (also in the diagrams)
• Disco also builds an Aggregator chosen from the task's strategy.

Memory:

• Dropped the reference to DistributedTrainer.memory as it no longer exist.

[JulienVig]

Thanks for the updates! I left a few minor comments, feel free to merge when they're addressed