[SPARK-58164][SQL] approx_top_k_estimate crashes or returns wrong results on sketches whose item type was widened by type coercion

[jiwen624]

What changes were proposed in this pull request?

approx_top_k_estimate picked its sketch deserialization SerDe from the static item type in the sketch-state struct (field 2). A UNION ALL of two sketch columns with coercible item types (e.g. INT + BIGINT) widens that static type while the serialized bytes and the per-row DDL (field 3) keep the original type. Estimation then decoded the bytes with the wrong SerDe.

This PR makes estimate deserialize using the true item type recorded per-row in the state DDL (recovering collation from the static type, exactly as combine already does), then widens each decoded item to the declared output type via a lossless cast so results still match the statically-declared schema.

Why are the changes needed?

A query like:

SELECT approx_top_k_estimate(approx_top_k_combine(sketch, 5))
FROM (SELECT approx_top_k_accumulate(c) AS sketch, 'int' AS tag
        FROM VALUES (1), (1), (2) AS t(c)
      UNION ALL
      SELECT approx_top_k_accumulate(c) AS sketch, 'long' AS tag
        FROM VALUES (CAST(9 AS BIGINT)) AS t(c))
WHERE tag = 'int';

fails at runtime with SketchesArgumentException: Bounds Violation, e.g.,

26/07/16 22:29:33 ERROR Executor: Exception in task 0.0 in stage 2.0 (TID 3)
org.apache.datasketches.common.SketchesArgumentException: Bounds Violation: reqOffset: 0, reqLength: 16, (reqOff + reqLen): 16, allocSize: 10
        at org.apache.datasketches.common.Util.checkBounds(Util.java:659)

Or, depending on the widening it can also throw ClassCastException or, when byte sizes
happen to match (e.g. BIGINT→DOUBLE), silently return wrong results.

Does this PR introduce any user-facing change?

Yes. A query that may crash or return wrong results now gets fixed.

How was this patch tested?

Added UT.

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

Yes