kernels: nvfp4: support decomposed linear (LoRA outlier compensation)

[zhitwang17]

Support DecomposedLinear (LoRA outlier compensation) under NVFP4

Summary

Enables DecomposedLinear low-rank outlier compensation for NVFP4 low-precision
training. Quantizing weights to FP4 corrupts a small set of outlier values; a low-rank
U·Σ·V term (LoRA-style) recovers that lost signal at negligible cost. This path already
existed for MXFP4 but was never exercised end-to-end under NVFP4 or under TorchTitan's
meta-device model construction, where it crashed during conversion.

The dispatch/quantization wrappers are scheme-agnostic, so no kernel changes are required;
the fixes are confined to DecomposedLinear construction and its test coverage.

Changes

• Meta-device-safe from_linear: allocate u/v/sigma on the source weight's
  device/dtype (incl. meta), fixing a set_data crash during TorchTitan's meta-device
  conversion.
• lora_rank validation: require a positive multiple of 16 (the FP4 GEMM block size)
  at construction, replacing an opaque deep-kernel torch._check failure with a fast,
  actionable error.
• Tests: parametrize the suite over NVFP4 + MXFP4, add a meta-device build/swap test,
  and trim the CUDA SNR matrix to [128, 512] for cheaper CI.

Op-level test results

All green; no regressions in the shared dispatch / quantization paths.

Suite	Result
tests/unittest/nn/test_decomposed_linear.py (incl. CUDA SNR, NVFP4 + MXFP4)	94 passed
nvfp4/ + mxfp4/ op-level suites (dispatch guards, linear, quantization)	526 passed, 400 skipped

E2E validation (debug model)

Three-way 10,000-step comparison on gpt-oss-debug (~82.8M total / 47.3M active, MoE),
GBS=16, 2 × MI300X. NVFP4 recipes are identical except for the compensation knob
(lora_rank 0 vs 32). All runs were stable: monotonic loss, healthy grad norms, no NaNs.

Recipe	Final train loss (step 10k)	Δ vs BF16	Δ vs NVFP4 rank0
BF16 baseline	0.0727	—	—
NVFP4, compensation off (rank0)	0.0901	+0.0174	—
NVFP4 + DecomposedLinear (rank32)	0.0833	+0.0106	−0.0068

DecomposedLinear closes ~39% of the NVFP4↔BF16 final-loss gap, confirming the
compensation is both correct and effective under NVFP4.