feat(write): add DataFrame.writeParquet with ParquetWriteOptions

[andygrove]

Which issue does this PR close?

• Closes #.

Rationale for this change

DataFusion's DataFrame::write_parquet is the natural sink for transformed data, but today the Java bindings have no way to write results back to Parquet — collect() into JVM-side Arrow batches is the only option. This blocks ETL-shaped workloads where read → transform → write needs to stay in native code.

What changes are included in this PR?

• New org.apache.datafusion.ParquetWriteOptions (fluent setters for compression and singleFileOutput), shaped to mirror the existing ParquetReadOptions.
• DataFrame.writeParquet(String) and DataFrame.writeParquet(String, ParquetWriteOptions) overloads. Both retain the DataFrame (clone on the Rust side), matching the existing count() / show() pattern; the receiver stays usable and must still be closed.
• One new JNI function Java_org_apache_datafusion_DataFrame_writeParquetWithOptions in native/src/lib.rs. Compression strings (e.g. "zstd(3)", "snappy", "uncompressed") are passed verbatim to DataFusion; invalid values surface as RuntimeException at write time.

Deliberately not included this round:

• An overwriteMode knob / Java InsertOp enum — DataFusion 53.1.0's write_parquet only implements Append (Overwrite and Replace raise "not implemented"). Re-adding the knob once upstream support lands is a non-breaking addition.
• partition_by, sort_by, row-group / page-size / dictionary / statistics / bloom-filter knobs — out of scope for the first cut.
• CSV / JSON / Avro write outputs and any shared WriteOptions base.

Are these changes tested?

Yes. Two test files added:

• ParquetWriteOptionsTest — pure-Java unit tests for defaults and fluent setter behavior.
• DataFrameWriteParquetTest — four integration tests guarded by Assumptions.assumeTrue(Files.exists(lineitem)) so they skip cleanly when tpch-data/ is absent:
  • Round-trip row count via the no-options overload (multi-file output).
  • singleFileOutput(true) produces a regular file at the supplied path.
  • compression("zstd(3)") writes and reads back with row count preserved.
  • DataFrame remains usable after writeParquet (pins the retain semantics).

make test shows 42 tests, 0 failures. cargo clippy --all-targets --workspace -- -D warnings is clean.

Are there any user-facing changes?

Yes — adds new public API (DataFrame.writeParquet overloads + ParquetWriteOptions). No removals or breaking changes.