Add adapter-only checkpointing to Megatron-LM

Parent initiative: #7267

## Work

Add adapter-only checkpoint save and load support to `megatron.training.checkpointing`.

## Done when

- Megatron-LM can save adapter weights without writing a second copy of the frozen base-model weights.
- Megatron-LM can load an adapter-only checkpoint onto the matching base model.
- The checkpoint contents remain compatible with the Megatron-Bridge PEFT workflow.
- Resuming PEFT training restores the adapter state required to continue training.
- Tests cover adapter-only save, load, and resume behavior.

Start in megatron.training.checkpointing and trace the existing save, load, and resume paths for PEFT training. Compare the checkpoint expectations of the Megatron-Bridge PEFT workflow, then add tests for adapter-only save, load onto a matching base model, and resume; done means frozen base weights are not duplicated and adapter state is restored compatibly.

Written by the indexing model from the issue text.