Backport v1 MFSDP full-iteration CUDA graph safety

[lhb8125]

Summary

• preserve v1 Megatron-FSDP gradient references required by full-iteration CUDA graph replay
• disable the optimizer-step parameter all-gather that crosses the full-iteration capture boundary
• preserve non-FSDP-unit parameter bucket storage across AllGatherPipeline reset
• keep preserved buckets eligible for a fresh all-gather before reuse

This is a focused backport of behavior already merged on dev through NVIDIA#4663 and NVIDIA#4717. It is independent from the FSDP v2 stable-storage and integration PRs.

The first commit is mechanical formatting required by the current formatter; the second commit contains the behavior change.

Validation

• Python 3.12 py_compile on all changed files
• CHECK_ONLY=true BASE_REF=mfsdp_refactor bash tools/autoformat.sh
• isolated v1-on-v1 baseline 20260703-014718-6375 / 2265910 at 337d07a35 completed 30/30 finite iterations plus 32 validation and 32 test iterations; warmup-8 mean/median were 1032.31/1058.90 TFLOP/s/GPU, with peak allocated/reserved/device-used 90,809/117,334/127,586 MB
• first required-set v1-on-v2 run 20260703-013644-1ba9 / 2265881 completed the same workload at 1027.27/1051.95 TFLOP/s/GPU, with peak memory 89,885/110,590/120,842 MB; final loss/grad norm were finite and aligned
• final head d7177a1aa v1 control 20260703-020830-a4a1 / 2266229 completed 30/30 finite iterations plus 32 validation and 32 test iterations; peak memory was 89,885/119,370/129,622 MB
• the final run had a transient slowdown in iterations 9-17; its stable tail (iterations 18-30) was 1026.50/1051.00 TFLOP/s/GPU, matching the first same-node control

[lhb8125]

Closing this PR because it is not a dependency of the FSDP v2 integration. The validated v1 baseline already contains upstream NVIDIA#4663 and NVIDIA#4717; relative to that baseline, NVIDIA#5636 is the only additional fix required for the target v1 full-iteration CUDA graph configuration. We will keep this branch/commit only as a temporary experiment dependency when running v1 full-CG on the older mfsdp_refactor base, until that base is updated to current dev or otherwise includes NVIDIA#4663 and NVIDIA#4717.