[Common] Ensure quantization kernels handle noop properly

[kainzhong]

Description

Fixes a few places where quantization treats the noop flag wrong.

Type of change

• Documentation change (change only to the documentation, either a fix or a new content)
• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Infra/Build change
• Code refactoring

Changes

Please list the changes introduced in this PR:

• MXFP8 & Group MXFP8 now ignore noop with IS_DBIAS since the reduction kernel will run unconditionally, so if we skip quantization with IS_DBIAS the results would be wrong.
• MXFP8's zero_scales_kernel now will ignore noop using the same condition as how the quantization ignores it (otherwise if we skip zeroing but not skip quantization the result could be wrong)
• Blockwise FP8's group_block_scaled_1d_tma_kernel and group_block_scaled_2d_tma_kernel now ignore the noop flag when dbias workspace is nullptr (I suppose this means IS_DBIAS) for the same reason
• Pass noop flags to FP8 quantization
• Add tests to check if noop flag is honored

Checklist:

• I have read and followed the contributing guidelines
• The functionality is complete
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[greptile-apps]

Greptile Summary

The PR aligns noop handling across FP8, MXFP8, grouped MXFP8, and blockwise FP8 quantization.

• Passes the noop flag into the aligned FP8 quantization kernel.
• Keeps quantization, scale zeroing, and unconditional DBIAS reduction behavior consistent.
• Adds CUDA-graph tests covering noop behavior across quantization formats, output layouts, and shapes.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains; the MXFP8 DBIAS path now passes a null noop pointer consistently to both quantization and padded-scale zeroing, so the previously reported stale-scale mismatch is fixed.

Important Files Changed

Filename	Overview
transformer_engine/common/cast/mxfp8/quantize_mxfp8.cuh	Makes DBIAS and activation variants consistently ignore noop in both quantization and padded-scale zeroing, resolving the previously reported stale-scale failure.
transformer_engine/common/cast/mxfp8/group_quantize_mxfp8.cuh	Prevents DBIAS-only grouped MXFP8 quantization from skipping workspace production required by unconditional reduction.
transformer_engine/common/cast/fp8_blockwise/group_quantize_fp8_blockwise.cuh	Restricts noop early returns to launches without a DBIAS workspace so reduction inputs are always produced.
transformer_engine/common/cast/fp8/quantize_fp8.cuh	Propagates noop into the aligned one-dimensional FP8 kernel and skips its output updates when requested.
tests/pytorch/test_quantized_tensor.py	Adds CUDA-graph coverage confirming noop preserves quantized data and scale buffers across supported formats and layouts.

_{Reviews (3): Last reviewed commit: "[Common] Ensure quantization kernels han..." | Re-trigger Greptile}