fix(cuda-graphs): preserve nested static input buffers

[FDecaYed]

What

Make full-CUDA-graph input refresh recursively reuse compatible static tensors inside tuples, lists, and dictionaries. Scalars and incompatible entries are replaced, while reusable tensors keep their object identity and storage.

This helper fix is self-contained and is intentionally detached from the DSv4 feature merge.

Relationship to #5795

• This PR is not a dependency or child of [Tracking] DeepSeek-v4 HybridModel merge plan #5795.
• The frozen HybridModel-only DSv4 recut uses Transformer Engine CUDA graphs, not the full-iteration CUDA-graph path.
• The implementation is therefore excluded from the updated [Tracking] DeepSeek-v4 HybridModel merge plan #5795 aggregate and can be reviewed and merged independently.

Provenance

• Reconstructs the focused nested-buffer correction from Enable Deepseek-v4 hybrid_model in dev branch Part (2/N) #5485.
• Original implementation by the Enable Deepseek-v4 hybrid_model in dev branch Part (2/N) #5485 authors; DSv4-era adaptation by @hxbai.
• No GPTModel-specific integration is included.

Testing

• uv run isort --check-only on both changed Python files
• uv run ruff check on both changed Python files
• uv run python -m compileall -q on both changed Python files
• Added a CPU-only nested tuple/list/dict regression test; full execution is left to CI because the local workspace lacks PyTorch.

[copy-pr-bot]

Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually.

Contributors can view more details about this message here.

[FDecaYed]

/ok to test f2475b1