Read Iceberg v3 deletion vectors in DataFusion scans

[JanKaul]

Summary

Adds read support for Iceberg v3 deletion vectors (DVs) in datafusion_iceberg scans.

• Parse deletion-vector-v1 Puffin blobs into an in-memory roaring bitmap (iceberg-rust-spec), loaded eagerly via ranged object-store reads keyed by the referenced data-file path (iceberg-rust).
• Apply DVs during the scan with a custom IcebergDvExec physical node that filters rows by (data_file_path, absolute_row_position) — a per-file delete-index rather than a join.
• Derive the absolute row position from DataFusion's native Parquet RowNumber virtual column, so filtering is correct under row-group pruning, predicate pushdown, and inter-/intra-file repartitioning.

This implements the read side of the deletion vectors section of the Iceberg spec. Writing DVs is a separate follow-up.

Architecture

Puffin DV blob ─> load_deletion_vectors ─> HashMap<data_file_path, DeletionVector(roaring)>
                                                        │
data scan (+ __data_file_path, RowNumber virtual col) ─┴─> IcebergDvExec  ─> filtered rows
                                                             (probe bitmap by true row number)

• Data files with no DV take the existing scan path untouched — zero overhead.
• The DV index holds only the compact roaring bitmaps; the scan stays streaming and preserves ordering.
• Predicate pushdown is enabled on the DV data scan (safe: supports_filters_pushdown is Inexact, so a FilterExec is retained above the scan).

Validation

• cargo test -p datafusion_iceberg --lib — unit tests for IcebergDvExec (non-contiguous positions from pruned row groups, files split across out-of-order batches, interleaved files, column stripping) plus an in-crate real-ParquetSource scan test proving true row numbers survive row-group pruning and drive the filter end to end.
• cargo clippy -p datafusion_iceberg --lib --tests -- -D warnings, cargo fmt --all -- --check.

Relationship to #412

#412 reads v2 position-delete files (Parquet). Both features filter rows by (data_file_path, row_position) and share the RowNumber primitive introduced here. Once this merges, #412 can integrate by feeding v2 position-delete files into the same per-file bitmap index consumed by IcebergDvExec, unifying v2 and v3 on one operator. A detailed integration plan is posted on #412.