Correlation alignment replaces zero-variance columns with independent noise on each side

`_align_by_correlation` (`nltools/algorithms/alignment/procrustes.py`, `replace_zero_variance=True`) draws independent uniform noise for the reference and the target when it replaces zero-variance columns, so the matched column's assignment in `linear_sum_assignment` is arbitrary rather than deterministic across the pair. The float-conversion fix from #535 (H-06) makes the replacement take effect; the independence of the two draws is a pre-existing wart the S-H verification report and implementer both flagged as out of scope.

Options: draw one noise vector per zero-variance column and reuse it on both sides, or exclude those columns from the assignment and map them last. Needs a decision on the intended semantics; the two existing tests only assert shape and index validity.

Start in nltools/algorithms/alignment/procrustes.py at _align_by_correlation with replace_zero_variance=True, then inspect the two existing tests for shape and index validity. Resolve the intended semantics for zero-variance columns with the maintainers, implement the chosen behavior, and extend the tests to verify deterministic assignment across the pair.

Written by the indexing model from the issue text.