chore: audit array_intersect and expand SQL test coverage

[andygrove]

Which issue does this PR close?

Closes #.

Rationale for this change

ArrayIntersect was marked Incompatible(None) with no explanation and had minimal test coverage: one SQL file exercising only array<int>, plus three Scala assertions. The audit surfaced three concerns that are worth making explicit:

1. DataFusion's array_intersect builds its hash lookup from the shorter input and probes the longer one, so element order in the result follows the longer side. Spark follows the left side. Results therefore diverge when the right argument is longer than the left, but no test exercised this case and the Incompatible marker had no reason text, so users had to set allowIncompatible=true on faith.
2. Neither getIncompatibleReasons() nor getUnsupportedReasons() was overridden, so even once a reason is added to getSupportLevel it would not show up in the generated Compatibility Guide.
3. Spark 4.0 introduces collated StringType. DataFusion compares rows by bytes, so array_intersect on a column with non-UTF8_BINARY collation would silently return a wrong answer instead of falling back.

What changes are included in this PR?

• Serde (CometArrayIntersect): add a reason string describing the probe-order divergence and surface it through both getSupportLevel and getIncompatibleReasons. Return Unsupported when the element type (recursively) contains a StringType with non-default collation, and surface that reason through getUnsupportedReasons too.
• Shim (CometTypeShim): add hasNonDefaultStringCollation(dt) that walks nested ArrayType / MapType / StructType. The Spark 4.0 variant returns true for any StringType whose collationId != UTF8_BINARY_COLLATION_ID; the Spark 3.x variant is stubbed to false.
• SQL test (array_intersect.sql): extend from 4 int-only queries to ~20 queries covering boolean, tinyint, smallint, bigint, float, double (NaN / +-Infinity / -0), decimal, date, timestamp, string (incl. multibyte UTF-8), binary, nested array<array<int>>, duplicate elements, self-intersection, all-NULL and both-NULL arrays, CASE WHEN inputs, and a right-longer-than-left query wrapped in sort_array so the documented ordering divergence stays a stable assertion.
• Support doc: add audit sub-bullets under array_intersect in spark_expressions_support.md for Spark 3.4.3 / 3.5.8 / 4.0.1 dated 2026-04-24, noting the ordering incompatibility and the 4.0 collation fallback.
• Skill (audit-comet-expression): add a reminder to the skill instructions that whenever getSupportLevel returns a reason, the corresponding getIncompatibleReasons() / getUnsupportedReasons() override must also be added so the Compatibility Guide picks it up.

How are these changes tested?

CometSqlFileTestSuite [array_intersect] and CometArrayExpressionSuite (39 tests) both pass locally. make succeeds on the Spark 3.4, 3.5, and 4.0 profiles, and cargo clippy --all-targets --workspace -- -D warnings is clean.