[SPARK-58095][SQL] approx_top_k_combine throws scala.MatchError on empty input

[jiwen624]

What changes were proposed in this pull request?

Fix the MatchError on empty input with changes below:

• Fall back to the statically-known item data type carried by the state struct (uncheckedItemDataType) when no input sketch resolved it, in both eval and serialize.
• When combine size is unspecified and no input was consumed, maxItemsTracked is still the internal VOID sentinel (-1); resolve it to the default in eval only (a partial buffer must keep VOID for the final merge), so the empty output sketch carries a valid size that downstream functions such as approx_top_k_estimate accept.

Why are the changes needed?

The query aborts with an internal error instead of returning an empty result, unlike the sibling functions (approx_top_k, approx_top_k_accumulate, approx_top_k_estimate) which handle empty input correctly:

SELECT approx_top_k_combine(sketch, 5)
FROM (SELECT approx_top_k_accumulate(expr) AS sketch
      FROM VALUES (1), (2) AS t(expr))
WHERE false;
-- before: org.apache.spark.SparkException ... scala.MatchError: null
-- after:  returns a single empty sketch (estimates to [])

Does this PR introduce any user-facing change?

Yes. Fixed an query failing issue when the input is empty.

How was this patch tested?

Added UT case

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

Yes.

[uros-b]

cc @cloud-fan here regarding https://github.com/apache/spark/pull/57375/changes#r3621607120

[jiwen624]

@uros-b @cloud-fan ah it seems I forgot to move it out of draft. I see there is a problematic scenario covered by this and currently not by the one above. I’m fine with either addressing it separately here or combining the fixes.