Merge NV 4.2.1 to SYCL-TLA Main

[anamikac-intel]

No description provided.

[Copilot]

Pull Request Overview

This PR merges NVIDIA CUTLASS version 4.2.1 into the SYCL-TLA main branch. The primary purpose is to update the version number from 4.1.0 to 4.2.1 and incorporate upstream changes from the NVIDIA CUTLASS 4.2.1 release.

Key changes include:

• Version bump from 4.1.0 to 4.2.1 across Python and C++ components
• Addition of SM100/SM103/SM120/SM121 architecture support and related utilities
• Bug fixes in existing code (matrix operations, cross product calculation)
• Renaming of "BlockScaled" terminology to "Blockwise" for FP8 kernels
• Addition of new Python test utilities for SM100 shapes and instantiation levels

Reviewed Changes

Copilot reviewed 73 out of 73 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
test/unit/gemm/device/gemm_testbed_3x_ptr_array.hpp	Refactored alignment calculation to handle ElementA and ElementB separately
test/python/cutlass/interface/gemm_interface.py	Updated compute capability checks and schedule validation
python/setup_pycute.py	Version bump to 4.2.1
python/cutlass_library/sm90_utils.py	Added blockwise schedule support
python/cutlass_library/sm100_utils.py	New file: SM100 kernel generation utilities
python/cutlass_library/sm100_shapes.py	New file: SM100 MMA and cluster shapes
python/cutlass_library/manifest.py	Renamed method for broader architecture support
python/cutlass_library/library.py	Added new kernel schedule types and SM100 support
python/cutlass_library/gemm_operation.py	Refactored SM103 FP4 ultra kernel schedule checks
python/cutlass_library/generator.py	Replaced manual instantiation with SM100 utilities
include/cutlass/matrix.h	Fixed unary negation operator bugs
include/cutlass/version.h	Version bump to 4.2.1

Comments suppressed due to low confidence (5)

test/unit/gemm/device/gemm_testbed_3x_ptr_array.hpp:1

• Trailing whitespace should be removed.
  media/docs/cpp/profiler.md:1
• Corrected spelling of 'defination' to 'definition'.
  media/docs/cpp/pipeline.md:1
• This correction appears to fix an error where "consumer threads" was changed to "producer threads". The producer_acquire operation should block producer threads, not consumer threads, so this is a correct fix.
  media/docs/cpp/cute/02_layout_algebra.md:1
• The stride calculation was corrected from (3*w,6*x,2*x,2*z) to (72*w,24*x,4*y,2*z). The original had an error using 2*x instead of the correct 4*y.
  media/docs/cpp/cute/02_layout_algebra.md:1
• The composition calculation was corrected from (3*w,3*x,y,z) to (9*w,3*x,y,z). The first stride should be 9*w not 3*w.