feat(scan): [3/N] read and apply V3 deletion vectors

[mbutrovich]

Which issue does this PR close?

• Closes [EPIC] Deletion vector read support (deletion-vector-v1 Puffin blobs) #2792.

What changes are included in this PR?

Draft PR, stacked on #2868. This is Task 3 of the deletion-vector read epic: reading and applying V3 deletion vectors during a scan.

• CachingDeleteFileLoader reads a deletion vector's deletion-vector-v1 blob directly from its Puffin file by byte range (content_offset / content_size_in_bytes), decrypting first via EncryptedInputFile when the entry carries key metadata.
• After decoding, the blob's cardinality is checked against the manifest entry's record_count, matching the same check Iceberg-Java does in BitmapPositionDeleteIndex.deserializeBitmap. Mismatch is a DataInvalid error, not a silent wrong result.
• DeleteFileIndex now indexes deletion vectors by referenced_data_file and applies them ahead of position delete files, since a DV supersedes any position deletes for the same data file per spec.
• A few spec invariants are now enforced as errors instead of silently mishandled:
  • a PositionDeletes entry with content_offset set but no referenced_data_file
  • two deletion vectors referencing the same data file (at most one is allowed per data file per snapshot)
  • a deletion vector whose partition or sequence number disagrees with the data file it claims to reference
• FileScanTaskDeleteFile gained a record_count field, populated from the manifest entry, used for the cardinality check above.
• DeleteVector::deserialize (from feat(delete-vector): [1/N] decode deletion-vector-v1 puffin blobs #2866) is now actually wired in; removed its #[allow(dead_code)].

No data-file read path changes outside of deletion vectors. Existing position delete and equality delete handling is unchanged.

Are these changes tested?

Yes.

• Unit tests in delete_file_index.rs for: DV supersedes partition-scoped and path-scoped position deletes, DV coexists with equality deletes, rejecting a DV missing referenced_data_file, rejecting duplicate DVs for one data file, rejecting a DV with a mismatched partition or stale sequence number.
• Unit tests in caching_delete_file_loader.rs for: reading and applying a DV, reading an encrypted DV, rejecting a cardinality mismatch, and the coordinate/cardinality validation helpers directly.
• End-to-end tests in positional_deletes.rs reading a real Parquet data file with a DV applied through ArrowReader, including the cardinality-mismatch failure path through the full read.

AI Disclosure

Developed with the help of Claude Code, but I understand and support these changes.