fix: clean up CometCast support-level reporting (#4501)

[andygrove]

Which issue does this PR close?

Closes #4501.

Rationale for this change

#4501 collected support-level and serde consistency follow-ups deferred from the cast expression audit in #4493. While working through them, several of the issue's line references turned out to be stale against current main, so the fixes here reflect the current state of CometCast.scala and GenerateDocs.scala.

Findings per item:

1. The issue noted that canCastFromFloat / canCastFromDouble / canCastFromDecimal ignore evalMode while every other arm of isSupported threads it through, and suggested the support level should reflect the eval mode. Investigation shows there is nothing to reflect: the native code does branch on eval mode for these narrowing casts (ANSI throws on overflow/NaN, LEGACY/TRY wrap), but Comet matches Spark in every mode, and CometCastSuite exercises them in LEGACY, ANSI, and TRY and passes. So there is no per-mode support-level divergence, and these three helpers (like canCastFromTimestamp) correctly take no evalMode. No code change needed.

2. For Cast, getIncompatibleReasons / getUnsupportedReasons are never consumed. cast.md is generated directly from the per-pair isSupported matrix (GenerateDocs.writeCastMatrixForMode, including per-pair note annotations), and EXPLAIN surfaces the per-pair reason via getSupportLevel. Cast is not in categoryPages, the only caller of those static methods. The generic overrides were dead, so they are removed in favor of the matrix as the single source of truth.

3. The array-of-binary Incompatible() branch described in the issue no longer exists; binary to string (including arrays) is currently Compatible(). The only remaining bare Incompatible() was the negative-scale decimal to string branch, which now carries a concise reason. The separate #4488 from_utf8_unchecked question for binary to string is a behavior change and is left to its own issue.

4. The literal short-circuit in getSupportLevel is correct: convert() folds literal casts via cast.eval() in Spark at planning time, so the cast never executes natively and the result matches Spark by definition. Only the comment was misleading; it is clarified.

What changes are included in this PR?

• Remove the unused getIncompatibleReasons / getUnsupportedReasons overrides from CometCast and document that the per-pair matrix is the canonical reason source.
• Fill in the empty Incompatible() reason for negative-scale decimal to string, sharing the string with the test via a constant to avoid drift.
• Clarify the literal short-circuit comment in getSupportLevel.

How are these changes tested?

Covered by existing CometCastSuite coverage, which exercises the float, double, and decimal cast paths in LEGACY, ANSI, and TRY modes. The negative-scale decimal to string test's support-level assertion is updated to reference the shared reason constant. No behavior changes, so no golden file or generated doc regeneration is required.

[andygrove]

@coderfender could you review?

[coderfender]

Thank you @andygrove