[CUDA] Add sm_121/Blackwell to known target

[Charlie-Tsai1123]

Summary

Add initial CUDA known target support for sm_121 / Blackwell NVIDIA GB10.

The CUDA execution limits are based on local cudaDeviceProp results from an sm_121 device. Existing NVIDIA MMA ops are reused as a conservative baseline until Blackwell-specific MMA intrinsics are modeled.

Related to #24477.
#24477 reports that IREE does not currently recognize newer Blackwell CUDA targets such as sm_120. This PR addresses the same target-enablement path for sm_121, which is the Blackwell target I can validate locally on NVIDIA GB10.
It intentionally does not add sm_120 support because I do not have sm_120 hardware to confirm the device limits or runtime behavior.

Testing

Tested locally on NVIDIA GB10 / sm_121. sm_121 requires PTX 8.8. Using +ptx88 compiles successfully.

Compiled and ran a local abs.mlir smoke test:

../iree-build/tools/iree-compile abs.mlir \
  --iree-hal-target-device=cuda \
  --iree-cuda-target=sm_121 \
  --iree-cuda-target-features=+ptx88 \
  -o abs_cuda.vmfb

../iree-build/tools/iree-run-module \
  --device=cuda \
  --module=abs_cuda.vmfb \
  --function=abs \
  --input=4xf32=-1,-2,3,-4

Results:

4xf32=1 2 3 4

Compiled and ran a local matmul.mlir smoke test:

../iree-build/tools/iree-compile matmul.mlir \
  --iree-hal-target-device=cuda \
  --iree-cuda-target=sm_121 \
  --iree-cuda-target-features=+ptx88 \
  -o matmul_cuda.vmfb

../iree-build/tools/iree-run-module \
  --device=cuda \
  --module=matmul_cuda.vmfb \
  --function=matmul \
  --input=128x256xf16=1 \
  --input=256x128xf16=1

Result: 128x128xf32 values are 256 as expected.

[AGindinson]

A LIT test would be nice to have, same as in PR #24525, once there's an alignment on the conflicts / order of merging these PRs.

[AGindinson]

LGTM, thanks!