feat: per-step inline HF export for Megatron checkpoints

[waleko]

Summary

• MegatronConfig gains an optional checkpoint: NotRequired[dict] passthrough, applied field-by-field onto megatron-bridge's CheckpointConfig in _create_checkpoint_config (hf_source_path defaults to the policy model name if unset).
• Lets you opt into megatron-bridge's inline HF export (also_save_hf_checkpoint=True) at every checkpoint save, not just via a separate end-of-training conversion step. Produces iter_*/hf/ — a directly-loadable HF safetensors + tokenizer tree — alongside the native Megatron checkpoint.
• DTensor backend is unaffected (this only touches megatron_cfg); it already has its own save_consolidated mechanism for full-model HF-shaped exports.

Usage

policy:
  megatron_cfg:
    checkpoint:
      also_save_hf_checkpoint: true
      hf_distributed_save: true       # spread the safetensors write across ranks
      ckpt_format: torch_dist         # required by also_save_hf_checkpoint
      non_persistent_ckpt_type: global  # required by also_save_hf_checkpoint

Test plan

• Compiles cleanly, no signature-breaking changes (both new params are optional/keyword)
• Verified live in a downstream fork's pipeline: also_save_hf_checkpoint=true produces a correct, directly-loadable iter_N/hf/ HF checkpoint at every save for a Megatron SFT run

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.