Consolidate smiles/smarts loaders further in benchmarks, ensure shuffling

[scal444]

No description provided.

[greptile-apps]

Greptile Summary

This PR completes the loader consolidation started earlier by ensuring every molecule loader always shuffles its output, eliminating file-order bias even when the input has fewer molecules than max_count. It also migrates butina_clustering_bench.py and cross_similarity_bench.py to use the shared load_smiles loader, and removes the duplicate local load_pickle/load_smiles/load_smarts implementations from substruct_bench.py.

• load_pickle, load_smiles, and load_sdf in bench_utils/loaders.py all gain an else: rng.shuffle(...) branch so benchmarks that consume a head slice always receive a representative, deterministically-seeded cross-section.
• butina_clustering_bench.py and cross_similarity_bench.py replace ad-hoc CSV/SMILES parsing with load_smiles(), gaining reservoir sampling and the new shuffle guarantee.
• ~130 lines of duplicate loader code are deleted from substruct_bench.py and superseded by the shared bench_utils implementations.

Confidence Score: 5/5

Safe to merge; changes are confined to benchmark utilities and eliminate a reproducible source of input-ordering bias.

All three loaders correctly handle both the sampling and the no-sampling paths, and both previously-flagged gaps (load_sdf and load_pickle missing shuffle-in-else) are addressed in this diff. The benchmark consumer files are straightforward call-site migrations with no logic changes.

No files require special attention.

Important Files Changed

Filename	Overview
benchmarks/bench_utils/loaders.py	All three loaders (load_pickle, load_smiles, load_sdf) now shuffle in the else-branch when sampling is not triggered, ensuring file-order bias is removed regardless of input size vs max_count.
benchmarks/butina_clustering_bench.py	Replaced manual head-slice CSV parsing with load_smiles(), gaining reservoir sampling and shuffle behaviour.
benchmarks/cross_similarity_bench.py	Replaced pandas CSV loading + manual parsing with load_smiles(); removes pandas dependency and gains shuffle; duplicate-padding fallback loop unchanged.
benchmarks/substruct_bench.py	Removed local load_pickle, load_smiles, load_smarts duplicates and replaced with imports from bench_utils; no logic changes to benchmark code itself.

_{Reviews (2): Last reviewed commit: "Shuffle sdf and pickle too" | Re-trigger Greptile}