Require CUDA 12.9 for SM121 default architectures

[davidkny22]

Summary

CUTLASS currently includes 121 and 121a in its default architecture list for CUDA 12.8,
but NVCC 12.8.61 advertises neither compute_121 nor sm_121. NVCC 12.9.41 advertises both.
This empirically locates the SM121 compile-support boundary at CUDA 12.9.

This PR moves only those two entries to the existing CUDA 12.9 block beside 121f.

Mechanism

CUTLASS_NVCC_ARCHS defaults to CUTLASS_NVCC_ARCHS_SUPPORTED, and the enabled entries are
translated to CMake CUDA_ARCHITECTURES. The current 12.8 block therefore makes a fresh default
build request two targets that its active compiler does not advertise. CUTLASS's architecture
header already requires CUDA 12.9 for CUTLASS_ARCH_MMA_SM121_SUPPORTED and
CUTLASS_ARCH_MMA_SM121A_ENABLED; this change aligns the default CMake policy with that gate.

Reproducer

For each installed toolkit, I ran a fresh native CMake configure and compared its default
CUTLASS_NVCC_ARCHS value with that compiler's own nvcc --list-gpu-arch output:

Toolkit	Before	After	SM121-family defaults
CUDA 12.8.61	2 unsupported of 17	0 unsupported of 15	2 → 0
CUDA 12.9.41	0 unsupported of 23	0 unsupported of 23	3 → 3
CUDA 13.0.48	0 unsupported of 23	0 unsupported of 23	3 → 3

Every non-moved default remains in the same order and maps to a target advertised by the
matching compiler. A CUDA 12.8 CUTLASS_NVCC_ARCHS=121 request still configures, remains cached
as 121, and is warned as unsupported.

Testing done

Fresh configure-only checks completed on Windows build 26200 with CMake 4.3.4, Ninja 1.13.0,
MSVC 19.44, and CUDA 12.8.61/12.9.41/13.0.48. The configures used
--allow-unsupported-compiler. The host GPU was an RTX 4070 Laptop GPU (SM89), driver 595.79;
no GPU execution was involved. Reconfiguring the pre-change CUDA 12.8 cache retained its old
list and warned about 121;121a.

CUTLASS has no in-tree test harness for the root version-to-default-architecture policy. I did
not add a parser or new test framework for this two-token policy change.

Limitations

• No SM121 runtime behavior was tested; this is a compiler-target policy change.
• The change affects fresh default lists. Existing configured build directories retain their
  cached CUTLASS_NVCC_ARCHS value until that cache is cleared or updated.
• The explicit CUDA 12.8 121 override was checked, but explicit overrides were not tested
  exhaustively.
• Validation covers NVCC 12.8.61, 12.9.41, and 13.0.48; it does not cover Clang-CUDA or other
  toolkit builds.

cc @Junkai-Wu