Mark non-final microbatches for MFSDP v2 hybrid data parallelism

[wujingyue]

Problem

MFSDP v2's hybrid data parallelism drops accumulated gradients: with multiple
microbatches, only the last microbatch's gradient reaches the optimizer.
reduce_partial_gradients uses context.is_last_microbatch to defer outer-DP
reduction, but the adapter inherited a no-op no_sync, leaving that flag true
for every backward. Finalizing the accumulation buffer on every backward caused
the next microbatch to clear it.

Changes

• Enter microbatch(is_last=False) from FullyShardedDataParallelV2.no_sync.
  Install the v2 no_sync_func regardless of overlap_grad_reduce, because v2
  reduces gradients during backward in either mode.
• Require the is_last_microbatch keyword at the gradient-reduction call site.
• Add accumulation regression coverage and three functional cases with new
  H100 dev golden values for NGC 26.08: baseline, HSDP, and HFSDP.
• Explicitly set expert_outer_dp_sharding_strategy='optim' in the HFSDP
  recipe and no_shard in the baseline and HSDP recipes. This PR preserves
  the current config and CLI defaults and does not depend on a default-inheritance change.

Validation

Rebased onto main at f6c33bde4; no merge commit.

• The eight-H100 hybrid unit suite passed before the default-inheritance
  change was split into a separate PR. This PR retains the
  gradient-accumulation regression coverage.
• Black, isort, and Ruff checks passed for the changed Core/test files.
• All three functional cases passed in independent fresh 50-iteration runs on
  8 x H100 with NGC 26.08: deepseek_proxy_mfsdp_v2_ep2_mb2,
  deepseek_proxy_mfsdp_v2_ep2_hsdp_mb2, and deepseek_proxy_mfsdp_v2_ep2_hfsdp_mb2.
  Each passed exact loss and approximate memory checks against its new goldens.
  Iteration-1 gradient norm was 10.681 for all three.
  The recipes now explicitly select the same effective expert outer strategies
  used in those runs; training was not rerun after that mechanical change.

Exact loss checks remain enabled. Approximate loss tolerance alone does not
reliably detect the original accumulation bug, so these cases must not adopt
NON_DETERMINSTIC_RESULTS=1. Each case checks its own golden; the functional
harness does not compare the three configurations directly.

[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.

[wujingyue]

is_last_microbatch is important enough to be required.

[shjwudp]

Neat!

[nemo-automation-bot]

🔄 Merge queue validation started!

You can track the progress here: https://github.com/NVIDIA/Megatron-LM/actions/runs/34711461179