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]

Confidence Score: 5/5

The PR appears safe to merge, with no actionable correctness, security, or repository-rule violations identified.

Summary

• Adds specialized tensor-kernel launch configurations for query sizes up to 8, 16, and 32 rows.
• Corrects the shared-memory output stride to use the tile’s X dimension.
• Adds boundary-shape and empty-fingerprint tests for the optimized path.
• Adds a compile-time check that the reused shared-memory allocation can contain each output tile.

_{Reviews (2) · Last reviewed commit: "Guard tensor tile shared-memory capacity"}

[evasnow1992]

Changes look good to me. One comment on whether it would be worth adding back a tile-size check in the new kernel launcher.