perf: vectorise the GainLoss decomposition and the CustomerDecisionHierarchy Yule's Q matrix #611

What: two analysis classes do per-row or per-pair Python work that is slow at retail scale. Both predate the current branch.

• GainLoss: the six-way decomposition calls process_customer_group once per customer through DataFrame.apply(axis=1). With millions of customers that is millions of Python calls. The same branch logic can be written once with np.select/np.where over the diff columns. process_customer_group is public, so keep it (or deprecate it) and add a test that the vectorised result matches it.
• CustomerDecisionHierarchy: the Yule's Q distance matrix is built in a Python loop over every product pair, calling .toarray() on a dense customer vector each time. That is O(products²) iterations over arrays as long as the customer count, so 1,000 products × 1M customers means about 500k iterations over 1M-length arrays. Sparse M @ M.T plus the row sums give the a/b/c/d counts for all pairs in one pass.

Both go against the CLAUDE.md rule to use vectorised operations instead of loops or .apply().

Tests: for each, assert the vectorised output equals the current output on the existing fixtures plus one randomised case.