fix(lumina): reject non-finite vector elements when building an index

[jackylee-ch]

extract_vectors_from_batches validated a null _ROW_ID, row-id contiguity, a null vector row, the dimension and a null element — but not finiteness, so a NaN or ±inf embedding was indexed silently. The everyday cause is a cosine normalisation dividing a zero vector by its zero norm.

Java rejects these per element in LuminaVectorGlobalIndexWriter#checkFinite, covered by testNanInVectorRejected and testInfinityInVectorRejected. The other Rust vector backend rejects them too, in paimon_vindex_core's validate_finite_values on both the write and the query side — so switching index.type between the two backends turned a loud build failure into a silent one.

The message carries the row id and the position inside the vector, as Java's does: a rejected build is only actionable if the offending row can be found.

[JingsongLi]

Reviewed head 8f21ee4. Requirement fit: supported; implementation: no blocking findings. The Lumina build path reads Arrow List and FixedSizeList through this shared element loop, and previously forwarded NaN/infinity to native index construction; rejecting the specific row ID and vector offset before commit makes that invalid input diagnosable and aligns with the existing vindex behavior. I ran both new targeted tests in an isolated build: test_extract_vectors_rejects_non_finite_element (NaN and ±infinity) and test_extract_vectors_fixed_size_list_rejects_non_finite_element; both passed. The check covers each non-null element without altering null/dimension/row-ID validation.