Further support for Blackwell and L-class GPUs

[scal444]

cc 89 (L40s) were left out of the fat build. Originally they had been supported by 75 PTX, but we removed it at some point so these were not working.

BMMA enabled for CC 10 and 12, I checked that it works.

[greptile-apps]

Greptile Summary

This PR adds 89-real to the fat-build arch list to restore L-class GPU (L40S/CC 89) support, enables BMMA tensor ops at runtime and compile time for Blackwell (sm_100, sm_120), removes the __CUDA_ARCH__ < 1000 upper bound in the similarity kernel guards, and fixes the getMaxThreadsPerSM value for consumer Blackwell (sm_120: 1536, not 2048). Open concerns flagged in the previous review round — the minor == 0 guard being too narrow for future Blackwell steppings and the PTX forward-compat gap for CUDA 12.8-only builds — remain unaddressed.

Confidence Score: 4/5

Safe to merge for current hardware; two concerns from the prior review round (minor-version guard and CUDA 12.8 PTX gap) remain open but do not affect production sm_100/sm_120 hardware today.

No new P0/P1 issues found in this pass. The logic for all three changed files is correct for the currently shipping Blackwell SKUs. The two previously flagged concerns (minor==0 narrowness and PTX forward-compat gap for CUDA 12.8 builds) are unresolved but are speculative/forward-looking rather than current breakage, keeping the score at the P1-ceiling of 4.

cmake/cuda_targets.cmake and src/similarity_kernels.cu carry the open forward-compat concerns from the prior review round.

Important Files Changed

Filename	Overview
cmake/cuda_targets.cmake	Adds 89-real to fix L-class GPU support and conditionally appends 100-real (CUDA ≥ 12.8) and 120 with PTX (CUDA ≥ 12.9); also extends the cc loop for 100/120 preprocessor defines. PTX forward-compat gap for CUDA 12.8-only builds and the minor-version guard remain open from the previous review round.
src/similarity_kernels.cu	Extends supportsTensorOps to accept major == 10 and major == 12, adds compile-time guards for CC_100/CC_120, and removes the __CUDA_ARCH__ < 1000 upper bound in the Tanimoto/Cosine kernel preprocessor guards to enable BMMA on Blackwell.
src/substruct/substruct_kernels.cu	Correctly adds explicit sm_100 (2048 threads/SM) and sm_120 (1536 threads/SM) cases in getMaxThreadsPerSM before the sm >= 90 catch-all, fixing the previously incorrect 2048 value that would have been returned for consumer Blackwell.

_{Reviews (4): Last reviewed commit: "Remove the extra brace, I blame greptile" | Re-trigger Greptile}

[evasnow1992]

Changes look good to me. Thanks!