fix: correct array_append return type and mark as Compatible

[andygrove]

Which issue does this PR close?

Closes #.

Rationale for this change

CometArrayAppend had two bugs:

1. Runtime assertion failure for literal arrays. DataFusion's array_append always returns a list with nullable: true on the element field. But when the input array has non-null elements (e.g. array(1, 2, 3)), Spark's ArrayAppend.dataType returns ArrayType(IntegerType, containsNull = false). The serde code was passing expr.dataType as the "promised" return type to DataFusion, which caused a runtime assertion failure:

   Assertion failed: result_data_type == *expected_type: Function 'array_append' returned value of
   type 'List(Field { data_type: Int32, nullable: true })' while the following type was promised at
   planning time and expected: 'List(Field { data_type: Int32 })'
   

2. Incorrect Incompatible classification. CometArrayAppend was marked Incompatible(None) with no explanation, which disabled it by default. The CaseWhen(IsNotNull(arr), array_append(arr, elem), null) wrapper already handles the only genuine incompatibility (DataFusion's array_append does not preserve null top-level array rows on its own), so the expression matches Spark's behavior fully.

What changes are included in this PR?

• arrays.scala: Use ArrayType(elementType, containsNull = true) as the promised return type for array_append, matching what DataFusion actually returns. Change getSupportLevel from Incompatible to Compatible.
• array_append.sql: Remove spark.comet.expression.ArrayAppend.allowIncompatible=true (no longer needed now that it's Compatible). Add comment explaining why ArrayInsert.allowIncompatible=true is still needed on Spark 4.0 (where array_append is a RuntimeReplaceable that rewrites to array_insert(-1)).
• expressions.md: Mark ArrayAppend as Spark-compatible.

How are these changes tested?

Existing SQL file test expressions/array/array_append.sql covers column inputs, literal inputs, NULL arrays, and NULL elements across both dictionary and non-dictionary Parquet. The literal-only query that previously triggered the assertion failure now passes.

[comphead]

Thanks @andygrove