Remove offline training, refactor train.py and logging/checkpointing

[aliberts]

What this does

• ⚠️ Removes the offline training part from the train.py script: online training will be handled by the training scripts from [WIP] Fix SAC and port HIL SERL #644
• In consequence, .offline and .online are removed from TrainPipelineConfig. To set the number of offline training step, simply use --steps:

python lerobot/scripts/train.py \
- --offline.steps=200000
+ --steps=200000

• Adds wandb_utils.py and turns Logger into WandBLogger to remove responsibilities from this class so that it only manages wandb stuff.
• Replaces training_state serialization with torch.save/load to safetensors.save_file/load_file. We shouldn't use torch.load() for this and in fact it breaks in which breaks in 2.6 due to weights_only=True by default.

/checkpoints/005000
  ├── pretrained_model
- └── training_state.pth
+ └── training_state
+     ├── optimizer_param_groups.json
+     ├── optimizer_state.safetensors
+     ├── rng_state.safetensors
+     ├── scheduler_state.json
+     └── training_step.json

• Adds train_utils.py to handle training checkpoints logic (including training state).
• Cleans up functions related to rng and groups them together in random_utils.py.
• Save checkpoint before eval during training rather than after (safer in case eval crashes)
• Fixes logging where displayed values would only be the last one measured instead of the average over the steps from previous logging step.
• Changed the policies main forward() output format for clarity. It now returns a tuple[Tensor, dict | None] instead of just a dict, the first element being the loss:

- output_dict = policy.forward(batch)
- loss = output_dict["loss"]
+ loss, output_dict = policy.forward(batch)
loss.backward()

How it was tested

Adds the following tests:

• tests/test_schedulers.py
• tests/test_optimizers.py
• tests/test_train_utils.py
• tests/test_random_utils.py
• tests/test_io_utils.py

How to checkout & try? (for the reviewer)

Examples:

pytest -v \
    tests/test_schedulers.py \
    tests/test_optimizers.py \
    tests/test_train_utils.py \
    tests/test_random_utils.py \
    tests/test_io_utils.py

[Cadene]

Beautiful

Could you remove all appearance of ema?
There were added by default

[Cadene]

Suggested change

	return loss, None
	# no output_dict so returning None
	return loss, None