fix: codegen dispatcher returns NULL for invalid try_make_timestamp inputs (#4554)

[andygrove]

Which issue does this PR close?

Closes #4554.

Rationale for this change

try_make_timestamp rewrites to MakeTimestamp(failOnError = false), which Spark's doGenCode implements with a try { ... } catch (DateTimeException e) { ev.isNull = true; } block. With Comet's codegen dispatcher routing MakeTimestamp through Spark's own codegen, the kernel was honoring the input-null short-circuit but skipping the post-eval ev.isNull guard, so invalid date/time components (month=13, day=32, hour=25, ...) returned stale ev.value bytes instead of NULL.

The dispatcher's defaultBody conflated NullIntolerant's "null in → null out" guarantee with the converse "non-null in → non-null out", which MakeTimestamp(failOnError=false) does not satisfy. The same path also affects non-ANSI make_timestamp directly, since that is the same expression with the same flag.

What changes are included in this PR?

• CometBatchKernelCodegen.defaultBody: in the NullIntolerant && allNullIntolerant branch, when the bound expression is nullable, wrap the post-ev.code write in if (ev.isNull) setNull else write. Non-nullable roots keep the direct write (the existing optimization).
• CometCodegenSourceSuite: lock in the fix with a generated-source assertion that MakeTimestamp(failOnError=false) emits two setNull sites — input short-circuit + post-eval guard.
• make_timestamp.sql: add column and literal queries with invalid components covering month/day/hour/minute out-of-range. These ride the default failOnError=false path (ANSI off in CometSqlFileTestSuite).
• try_make_timestamp.sql (new, Spark 4.0+): direct coverage for try_make_timestamp with both column and literal invalid inputs, plus a sentinel valid query so the dispatcher must run natively.

How are these changes tested?

• CometCodegenSourceSuite (Spark 3.5): 51/51 pass; the new test fails on the unmodified codegen and passes with the fix.
• CometSqlFileTestSuite make_timestamp filter (Spark 3.5 and 4.0): all three fixtures pass; on 3.5 try_make_timestamp.sql is skipped via MinSparkVersion: 4.0.
• All datetime SQL fixtures (Spark 3.5): 89/89 pass.
• CometCodegenSuite, CometCodegenHOFSuite, CometTemporalExpressionSuite: pass.

[mbutrovich]

Thanks for the fix, @andygrove! I really like seeing these types of fixes to codegen since I'm sure it fixes more than just this one expression that you tried.