Gradient checkpointing

In a complete training YAML, set the layers to recompute during backward:

training:
  use_ckpting: [ffn, mha, lora]

Supported names are ffn, mha, and lora. Checkpointed modules must participate in gradient computation. This trades extra computation for lower activation memory during training; it is not an inference cache option.