feat: add blob v2 columns through merge_columns_df

[everySympathy]

Summary

Adds support for adding BlobV2 columns through merge_columns_df by allowing callers to mark selected new columns as BlobV2 payload columns.

Changes

• Adds a blob_columns parameter to daft_lance.merge_columns_df / merge_columns_from_df.
• Converts marked columns with lance.blob_array(...) before merging.
• Routes BlobV2 add-column requests through the existing keyed merge slow path.
• Adds test coverage that creates a Lance 2.2 dataset, adds a BlobV2 column, and verifies the stored payload through take_blobs.

Testing

• uv run --python /usr/bin/python3.11 pytest tests/io/lancedb/test_fast_path_merge.py::TestBasicCorrectness::test_fast_path_adds_blob_v2_column -v
  • Result: passed, EXIT:0

Additional investigation:

• Full test_fast_path_merge.py on this branch:
  • Result: 38 passed, then process exits 139
• Excluding the new BlobV2 add-column test on this branch:
  • Result: 37 passed, 1 deselected, then process exits 139
• Same full file on main/baseline:
  • Result: 37 passed, then process exits 139
• A standalone fast-path merge script also reproduces EXIT:139.
• The same standalone scenario forced through slow path exits 0.

Notes / Risks

The 139 is a pre-existing native teardown / heap corruption issue in the current fast-path merge tests, not introduced by this BlobV2 add-column change.

The BlobV2 add-column implementation intentionally bypasses fast path because the current raw LanceFileWriter + manually stitched FragmentMetadata fast path is the path associated with the native exit crash. This PR should either be submitted as draft with the known fast-path stability issue called out, or submitted after a separate tracking issue is filed for the existing fast-path crash.

Suggested tracking issue:

test_fast_path_merge.py fast path causes native heap corruption on interpreter exit