Expand cross-similarity benchmark to larger sizes and trim variants

[scal444]

Replace the 100/1000/2000 molecule scan with a 2k-32k sweep driven by a shared fingerprint set generated once up front. Drop the multiprocess rdkit and nvmolkit CPU-collect benchmarks, keeping only the serial rdkit baseline and the nvmolkit GPU-only path (the others were for reference/experiment). Cosine similarity is now opt-in via a --cosine flag so default runs only time Tanimoto.

[greptile-apps]

Greptile Summary

This PR refactors the cross-similarity benchmark to generate one shared fingerprint set upfront for up to 32k molecules, then slice it per size — avoiding repeated fingerprint computation. The multiprocess rdkit and nvmolkit CPU-collect paths are removed, cosine similarity is now opt-in via --cosine, and the previously flagged infinite-loop risk is addressed with an explicit empty-mol guard and ValueError.

Confidence Score: 5/5

Safe to merge — clean benchmark-only refactor with no runtime logic or data-path changes.

No P0 or P1 findings. The infinite-loop concern from the previous review is resolved. The runner.args.values mutation pattern is synchronous with respect to bench_func calls, so per-benchmark value counts behave as intended. All SIZES values are ≤ max_size, so tensor slices are always in bounds.

No files require special attention.

Important Files Changed

Filename	Overview
benchmarks/cross_similarity_bench.py	Benchmark refactored to generate one shared fingerprint set up front, sweep 2k–32k sizes, drop multiprocess/cpu-collect variants, and add --cosine opt-in; previous infinite-loop bug is fixed with a not mols guard.

_{Reviews (3): Last reviewed commit: "Make slow CPU path singlular" | Re-trigger Greptile}

[evasnow1992]

Changes look good to me. Thank you for cleaning up this benchmark script!