[SPARK-55500][SQL] Fix analyzer cycle between ApplyDefaultCollation, ExtractWindowExpressions and CollationTypeCasts

[ilicmarkodb]

What changes were proposed in this pull request?

Call CollationTypeCasts immediately to avoid cycle between ApplyDefaultCollation, ExtractWindowExpressions and CollationTypeCasts.

Example:

CREATE TABLE t (c1 STRING, c2 STRING);  -- c1 is UTF8_BINARY                                                                                                                                                                                   
CREATE TABLE t2 DEFAULT COLLATION UTF8_LCASE AS
  SELECT c1 = 'HELLO', ROW_NUMBER() OVER (PARTITION BY c1 ORDER BY c2) FROM t;

Analyzer runs rules in batches sequentially, and the rule order is:
ApplyDefaultCollation -> ExtractWindowExpressions -> CollationTypeCasts.

Iteration 1:

• ApplyDefaultCollation applies UTF8_LCASE collation to the literal 'HELLO'. Expression EqualTo (c1 = 'HELLO') is not resolved after this because c1 and the literal have different types.
• ExtractWindowExpressions tries to extract the window expressions, but it can't because it expects the whole projectList to be resolved (both EqualTo and ROW_NUMBER()). EqualTo is not resolved, so it can't apply the rule.
• CollationTypeCasts applies coercion rules, changing the type of the literal to the default StringType again (because that's the type of the column). Now EqualTo is resolved.

Iteration 2:

• ApplyDefaultCollation again applies UTF8_LCASE collation to the literal, because its type is default StringType again.
• ExtractWindowExpressions again can't extract the window expressions for the same reason as before.
• CollationTypeCasts applies the same coercion rules again, changing the type of the literal back to the default StringType.

This cycle continues, and the plan never gets resolved. By calling CollationTypeCasts right after this rule, we ensure that other rules like ExtractWindowExpressions that expect resolved expressions can be applied.

Why are the changes needed?

Bug fix.

Does this PR introduce any user-facing change?

No.

How was this patch tested?

New tests.

Was this patch authored or co-authored using generative AI tooling?

No.