[SPARK-48685][ML] Fix MinHashLSH to handle empty sparse vectors without crashing

[ngtanvan]

What changes were proposed in this pull request?

Fix MinHashLSHModel.hashFunction to return Array.empty[Vector] when encountering empty sparse vectors (0 non-zero entries) instead of throwing an unhandled IllegalArgumentException.

Why are the changes needed?

When pipelines combining CountVectorizer and MinHashLSH process empty sparse vectors (e.g. text containing terms outside vocabulary), MinHashLSH threw java.lang.IllegalArgumentException: requirement failed: Must have at least 1 non zero entry., crashing the entire Spark job.
Returning an empty vector array allows posexplode in approxSimilarityJoin to cleanly skip empty vectors without crashing.

Does this PR introduce any user-facing change?

No breaking change. It fixes an unexpected crash when transforming or joining empty vectors with MinHashLSH.

How was this patch tested?

Updated unit test MinHashLSHSuite ("hashFunction: empty vector") to verify returning empty vector array, and verified all 15 tests in MinHashLSHSuite pass cleanly.

[ngtanvan]

Hi @HyukjinKwon , @srowen

Could you please help take a look or trigger CI for this PR when you have a moment?
This PR fixes [SPARK-48685] where MinHashLSHModel.hashFunction throws an unhandled IllegalArgumentException when encountering empty sparse vectors. All 15 unit tests in MinHashLSHSuite have been updated and pass cleanly.
Thank you very much for your time and assistance!

[tdcmeehan]

Please remove the unnecessary whitespace only change.

[tdcmeehan]

Would it make sense to follow other Spark ML transformers and expose this through handleInvalid with error (default), skip, and keep? keep could produce the proposed empty hash array. This preserves existing behavior while letting users opt into tolerant handling, consistent with VectorSizeHint and VectorAssembler.

[ngtanvan]

Thank you @tdcmeehan for the great suggestion! Adding handleInvalid (error, skip, keep) makes a lot of sense to maintain backward compatibility while providing flexibility consistent with other Spark ML transformers.

I will work on updating the PR to implement HasHandleInvalid. Thanks again for the review!