[lohedges]

Closes #561.

_score_rdkit_mappings paired every match of molecule0 with every match of molecule1. It checked each resulting mapping for duplicates by scanning a list of dicts, so the cost grew quadratically with max_scoring_matches. Most matches are permutations of equivalent hydrogens, so drug-sized ligands could take minutes per pair at the default of 1000.

Changes, applied to both the main module and the Sandpit copy:

• Duplicate mappings are skipped using a set of NumPy-built keys, so only new mappings are converted, aligned and scored. Pairs are visited in the same order as before.
• Each molecule's coordinates are read once. rmsd_align now uses a NumPy Kabsch superposition of the mapped atoms rather than a Sire move().align() and per-atom property lookups. rmsd_flex_align still uses flexAlign.
• Each mapping is now aligned from the original coordinates. Previously each alignment started from the previous mapping's result, and from that start Sire's alignment sometimes ended up in a worse orientation without raising an error. For example, every other benzene → toluene mapping scored 1.64 Å instead of 0.12 Å. The best mapping was unaffected in every case tested, but the rest of the ranking depended on the order mappings were visited.

Pair (1000 matches)	Before	After
Factor Xa 4d → 4c	> 150 s	0.39 s
TYK2 ejm_44 → ejm_43	4.1 s	0.03 s
BACE lig_41 → lig_67	3.3 s	0.03 s

Checked against devel on the test-suite and OpenFE benchmark ligands, in both directions, with and without a prematch:

• rmsd_align and rmsd_flex_align give identical ranked mappings.
• Plain rmsd gives the same mappings and scores, but mappings with exactly equal scores can come out in a different order, because the RMSD is summed in a different order. The best mapping is unchanged.

The new test_score_rdkit_mappings compares the function with a brute-force reference: every unique mapping from every pair of matches, with prematch filtering, and each aligned score checked against Horn's quaternion method. It fails 6 of its 20 cases with the previous implementation.

• I confirm that I have merged the latest version of devel into this branch before issuing this pull request (e.g. by running git pull origin devel): [y]
• I confirm that I have permission to release this code under the GPL3 license: [y]