Optimize Tanimoto similarity for skinny queries

[scal444]

Optimizes small dimension queries by transposing the small dim and using optimal matrix dimensions. N=1 is still inefficient, but much less so.

[greptile-apps]

Greptile Summary

Optimizes Tanimoto cross-similarity when the first input is skinny by swapping the tensor-operation axes and transposing writes back into the public matrix layout.

• Adds specialized 8-, 16-, and 32-column tensor-kernel launch configurations.
• Corrects shared-memory result stride handling for asymmetric tiles.
• Adds CUDA tests covering tile thresholds, rectangular matrices, fingerprint widths, and empty fingerprints.

Confidence Score: 5/5

The PR appears safe to merge, with no actionable correctness, security, or build failures identified.

The asymmetric tile mappings, shared-memory indexing, transposed output storage, architecture-specific compilation paths, and tested boundary configurations preserve the existing similarity matrix and metric contracts.

Important Files Changed

Filename	Overview
src/similarity_kernels.cu	Adds asymmetric tensor-kernel launches and transposed output indexing while preserving the public query-by-database matrix layout.
src/load_store.cuh	Changes the shared-memory result stride to the tile’s X dimension, matching kernel readback for asymmetric tiles.
nvmolkit/tests/test_similarity.py	Adds reference-based coverage across skinny-query tile boundaries and verifies empty-fingerprint Tanimoto behavior.

_{Reviews (1): Last reviewed commit: "Optimize Tanimoto similarity for skinny ..." | Re-trigger Greptile}