fix: treat empty DataFrame writes as noop on immediate-commit path

[yuw1]

Summary

When a Spark DataFrame pipeline produces zero rows upstream, df.write.format("lance").save(uri) currently fails with IllegalArgumentException: fragments cannot be null or empty. This PR makes the immediate-commit path of LanceBatchWrite.commit() a no-op on empty fragments.

The staged-commit path (Spark v2 DataFrameWriter's CTAS-style API, i.e. df.writeTo("t").create()) already accepts empty fragments via Operation::Overwrite and is the canonical way to create an empty table — that path is intentionally left untouched.

Fix

Short-circuit LanceBatchWrite.commit() on the immediate-commit path:

// Spark sink contract: empty immediate append writes should be a noop (not raise).
// Lance-Java Append rejects empty fragments, so skip building the operation.
// StagedCommit goes through Overwrite which does not enforce this check, so we
// only short-circuit the non-staged path. Empty overwrites must still go through
// Overwrite.builder() so the table is truncated and OCC is enforced.
if (stagedCommit == null && !isOverwrite && fragments.isEmpty()) {
  return;
}

Tests

• LanceBatchWriteTest.testCommitEmptyFragmentsIsNoop — unit test reproducing the bug on the immediate-commit path.
• LanceBatchWriteTest.testCommitEmptyFragmentsStagedIsNoop — unit test guarding the staged-commit path (path-based staged create of an empty table).
• LanceBatchWriteTest.testCommitEmptyOverwriteTruncatesTable — regression guard for the empty overwrite truncate path.
• integration-tests/test_lance_spark.py::TestDMLInsert::test_write_empty_dataframe_is_noop — pyspark end-to-end test mirroring the original repro.
• integration-tests/test_lance_spark.py::TestDMLInsert::test_write_empty_dataframe_overwrite_clears_table — pyspark end-to-end test for the empty overwrite truncate (local filesystem backend).
• integration-tests/test_lance_spark.py::TestDMLInsert::test_append_empty_dataframe_to_existing_path_is_noop — pyspark end-to-end test exercising the immediate-append guard on an existing Lance path (seed → empty append → seed rows unchanged).

Validation:

• LanceBatchWriteTest: 586 / 586 tests pass on lance-spark-base_2.12.
• spotless:check / checkstyle:check clean.
• git diff --check clean.
• Manually reverted the fix to confirm testCommitEmptyFragmentsIsNoop fails with the same IllegalArgumentException (fragments cannot be null or empty) at LanceBatchWrite.commit → Append.<init> (the Preconditions.checkArgument(!isEmpty()) guard), matching the original bug report.

Diff

 integration-tests/test_lance_spark.py              |  67 +++++++++++
 .../org/lance/spark/write/LanceBatchWrite.java     |   9 ++
 .../org/lance/spark/write/LanceBatchWriteTest.java | 123 +++++++++++++++++++++
 3 files changed, 80 insertions(+), 7 deletions(-)

Closes #768

[lance-gatekeeper]

✅ Gate recommendation: approve.

The mode-specific empty-write behavior is now correct, and the added integration test exercises the immediate empty-append path against an existing dataset.