Fix FSDP hooks bypassed by TE fused experts

[lhb8125]

Summary

• Forward GroupedLinear pre-forward hooks through the TE fused expert implementation while preserving PyTorch with_kwargs signatures.
• Forward submodule post-forward hooks to the fused MLP output so Megatron-FSDP can attach its pre-backward parameter all-gather.

This is the mfsdp_refactor backport of the core post-forward fix in NVIDIA#5636, plus the pre-forward hook-signature correction needed by with_kwargs=True fine-grained hooks on this branch. The PR contains production code only.

The rewrite removes the previous HybridEP static-budget fallback, token/probability/gradient row trimming, padding metadata/cache/zeroing, CUDA-graph handle retention, and related documentation/tests. The target paged-stash recipes already set moe_expert_rank_capacity_factor: 1.2; the automatic fallback was inactive there and lacked a complete overflow contract.

Validation

• CHECK_ONLY=true BASE_REF=mfsdp_refactor bash tools/autoformat.sh: Black, isort, Pylint 10.00/10, and Ruff passed.
• Copyright, Python 3.12 py_compile, and git diff --check passed.
• Same-node 64-GPU GB200 A/B, TP1/PP1/EP64, MBS2/GBS512, seq 4096, paged stash, explicit capacity 1.2, activation offload, full-iteration CUDA graph; metrics skip 8 warmup iterations.

Path	Control job / median TFLOP/s/GPU	Minimal job / median TFLOP/s/GPU	Delta	Peak allocated delta	Reserved/device-used delta
FSDP v2	20260702-202730-fdfa / 1033.50	20260702-203255-400a / 1033.15	-0.034%	-988 MB	-60/-60 MB
FSDP v1 on v2 code	20260702-203814-177d / 1050.05	20260702-204330-f5ca / 1045.70	-0.41%	-989 MB	-4280/-4280 MB

All four jobs completed 16/16 iterations with finite loss/grad norm, zero skipped iterations, and zero NaN iterations. Final control/candidate losses differed by only 1e-5 on each path.

The PR remains draft while the reduced scope is reviewed.