feat: expose io_buffer_size in CompactionOptions

[aimanmalib]

Summary

CompactionOptions did not expose io_buffer_size, even though the scanner used during compaction supports it. Compaction builds its scan reader in prepare_reader, which only forwarded batch_size to the scanner — the io_buffer_size knob was never set.

This matters because a single batch larger than the I/O buffer size causes the scanner to deadlock (documented on Scanner::io_buffer_size), and since the backpressure warning was downgraded to debug, this deadlock is now silent at the default. Users had no way to raise the buffer to avoid it during compaction.

Resolves #4946.

Changes

• Add io_buffer_size: Option<u64> to CompactionOptions (with Default = None).
• Support the lance.compaction.io_buffer_size manifest config key in apply_dataset_config.
• Plumb the value through prepare_reader → scanner.io_buffer_size(...), mirroring the existing batch_size handling.
• Update the Python binding to keep parameter names consistent across languages:
  • parse_compaction_options accepts io_buffer_size
  • CompactionOptions TypedDict documents the field

Testing

• Extended test_from_dataset_config to assert the lance.compaction.io_buffer_size key round-trips.
• Added test_compact_with_io_buffer_size (parametrized over Legacy/Stable file versions) that runs compact_files with an explicit io_buffer_size and verifies the compaction succeeds and preserves all rows.
• cargo test -p lance --lib dataset::optimize::tests → 76 passed, 0 failed.
• cargo clippy -p lance --lib --tests → clean.
• cargo fmt --all -- --check → clean.

Notes

This is a non-breaking, additive change — the new field defaults to None, preserving existing behavior when unset.