Choose and apply the best reindexing by correlation

[tjlane]

! This PR was vibe-coded.

Stack

Part 3 of 11 for #31. Depends on #352. Merge bottom-up.

Next branch: codex/phase-alignment-04-reindex-hardening.

What this implements

Adds ReindexingCandidate, ReindexingResult, and reindex_by_correlation(). The public function gates on the ambiguity predicate, evaluates identity plus Gemmi twin-law representatives, applies the highest-scoring operation to a copy, and reports ranked candidates, runner-up correlation, and the score gap.

Amplitudes are squared, normalized by resolution, and scored with signed Pearson correlation. Proper reindexing laws only are considered; hand remains separate.

Tests

End-to-end tests cover nonmutation, correct and identity selections, operation direction, common-HKL matching, amplitude and intensity inputs, and rejection when the unit-cell metric has no alternative indexing.

Reviewer focus

• The convention that the returned operation is applied to the moving dataset to match the reference.
• Candidate enumeration as identity plus Gemmi indexing-coset representatives.
• The returned diagnostics and normalized-correlation statistic.

Related: #31 and #174.