feat: Support DELETE operation on Lance datasets

[fightBoxing]

Summary

Implements DELETE support for Lance datasets, addressing #8.

Inspired by the lance-spark DELETE implementation, this PR adds predicate-based row deletion using Lance SDK's native Dataset.delete(String predicate) method.

Changes

New Files

• LanceDeleteExecutor: Standalone executor class that wraps Lance SDK's Dataset.delete() method
  • Supports predicate-based deletion (SQL-like syntax)
  • Provides delete(), deleteAndCount(), and countRows() methods
  • Implements Closeable for proper resource management

Modified Files

• LanceSink: Enhanced to handle DELETE RowKind
  • Separate delete buffer for batch processing DELETE operations
  • Builds delete predicates from RowData fields automatically
  • Supports common data types: BOOLEAN, TINYINT, SMALLINT, INTEGER, BIGINT, FLOAT, DOUBLE, CHAR, VARCHAR
  • Properly handles NULL values in predicates
  • Flushes deletes before inserts at checkpoint boundaries
• LanceDynamicTableSink: Updated ChangelogMode to support DELETE when requested

How DELETE Works

When a RowData with RowKind.DELETE is received:

1. The row is buffered in a separate delete buffer
2. When the buffer is full (or at checkpoint/close), a predicate is built from all buffered rows
3. The predicate combines all field values with AND for each row, and OR between rows
4. Example: For rows (id=1, name='Alice') and (id=2, name='Bob'):

   (id = 1 AND name = 'Alice') OR (id = 2 AND name = 'Bob')

5. The predicate is executed via Dataset.delete(predicate)

Usage Example

Standalone DELETE:

try (LanceDeleteExecutor executor = new LanceDeleteExecutor("/path/to/dataset")) {
    executor.delete("id = 1");
    executor.delete("age > 30 AND status = 'inactive'");
}

Testing

• Added 15 unit tests covering:
  • LanceDeleteExecutor creation and validation
  • Predicate validation (null, empty, blank)
  • ChangelogMode support (INSERT-only and INSERT+DELETE)
  • LanceSink builder validation
  • Copy and summary methods

Closes #8