Add option to save only trainable (non-frozen) model state in CheckpointManager.save_model_checkpoint + fix syntax error

[guillaume-byte]

Description:
When fine-tuning only the head (most of the base model frozen), saving full checkpoints wastes a lot of disk space.
We should allow saving:

• the full model state (current behavior)
• only the non-frozen/trainable parameters (plus relevant buffers if needed), so head-only finetunes don’t store huge frozen backbones like openVLA 7B parameters.

Place to work:
weightslab/components/checkpoint_manager@save_model_checkpoint