[SQL] Support for COVAR_* and REGR_* aggregate functions - #6812

Clean addition of COVAR_POP / COVAR_SAMP / REGR_SXX / REGR_SYY (REGR_COUNT is handled by the existing SqlCountAggFunction path).

Nice touches:

• guardedSum replaces Calcite's internal 2-arg SUM(x, guard) with a CASE WHEN guard IS NOT NULL THEN x END wrapped in SUM. That's the right lowering — Calcite's SUM(x, y) is a private convention that never really worked here.
• Linear implementation when args are non-FP: 5-tuple accumulator (sumAB, sumA, sumB, n, weight-1), computed in post via covariance(). Non-linear fallback for FP uses a 4-tuple with MUL_WEIGHT and the new QuadSemigroup — symmetric with TripleSemigroup and correct (each field uses DefaultOptSemigroup).
• Correct NULL semantics: bothNonNull mask ignores pairs where either side is NULL; COVAR_SAMP with n=1 yields DIV_NULL(_, 0)=NULL; REGR_COUNT=0 group returns NULL for cp/sxx/syy and 0 for rc. All tested in CovarTests.
• Bonus: the doVariance cleanup drops a redundant sq.cast(typedZero.getType()) after the MUL that was already producing typedZero.getType(). Silent simplification.

Tests (CovarTests + testRegrValue + testWindowCovariance) cover global, grouped, filtered, empty-partition, mixed-NULL, integer, and double paths. Postgres cross-check is called out in the class doc.

LGTM.

Re-approving on the new commit: the covariance family (COVAR_POP/COVAR_SAMP/REGR_SXX/REGR_SYY) implementation looks solid — linear-path decomposition into (Sum(a*b), Sum(a), Sum(b), Count) with bothNonNull masking and DIV_NULL in the post step correctly yields SQL NULL semantics (empty group → NULL for all except REGR_COUNT which is 0; single row → COVAR_SAMP NULL). CovarTests exercises grouped/global/FILTER/DOUBLE paths and the empty-set edge case. New QuadSemigroup mirrors the existing pair/tri variants — clean.

The guardedSum refactor in expandCovariance (SUM(CASE WHEN guard IS NOT NULL THEN value END)) is the right fix for REGR_SLOPE/INTERCEPT/R2 expansions that Calcite otherwise emits as internal 2-arg SUM(x, guard).

One tiny typo in the new comment (inline).

This branch contains some adversarial tests that fail along with fixes for two of the issues that cause the failures. These two seem like real bugs. The other two issues claude found -- I'm less clear, will leave it to you to decide how much of this to adopt.

Rebased tip 4bf3d61e71 over v0.330.0. Delta vs prior approved tip abe6e08d:

• docs/sql/aggregates.md: expanded the "widen your type" note to call out quadratic-blowup aggregates (VAR_*, REGR_*), and lowercased two count(*) to COUNT(*) for casing consistency. Improves the docs.
• CovarTests.java: added 22 lines — a WHERE FALSE empty-group case in testCovarGlobal (COVAR_POP → NULL, REGR_COUNT → 0). Nice edge-case coverage.

Core aggregator logic (linear 5-tuple / non-linear 4-tuple + QuadSemigroup), guardedSum rewrite, and NULL semantics all identical to the approved version. Re-approving on the current base.