feat(java): expose ArrowArrayStream export on LanceScanner

[sezruby]

Summary

Add LanceScanner#exportArrowStream(ArrowArrayStream) — a public wrapper around the existing private native openStream(long) JNI call. Lets callers populate a stream they allocated themselves instead of going through scanBatches(), which immediately imports the result into a Java ArrowReader backed by Lance's BufferAllocator.

Why

Consumers loaded under a different classloader and/or pinned to a different Apache Arrow version cannot safely share org.apache.arrow.vector.* classes with Lance — the JVM treats them as distinct types even when the bytecode is identical. The C Data Interface struct is stable across Arrow versions, so handing the C struct's memory address across the boundary is the only correct integration shape.

A concrete consumer is the gluten-spark / Velox integration tracked at apache/gluten#12263. gluten-spark builds against Arrow 15 (matching what Spark 3.5 ships and Velox uses); Lance Java SDK is on Arrow 18. With this method, gluten can:

try (ArrowArrayStream stream = ArrowArrayStream.allocateNew(glutenAllocator)) {
  scanner.exportArrowStream(stream);
  try (ArrowReader reader = Data.importArrayStream(glutenAllocator, stream)) {
    // import each batch into Velox via gluten's own Arrow 15 stack
  }
}

…where glutenAllocator is a Spark-task-managed BufferAllocator (ArrowReservationListener plumbing for memory accounting). Lance never sees Java Arrow on this side; ownership stays with the caller via the C Data Interface release callback.

What changed

• LanceScanner#exportArrowStream(ArrowArrayStream) — new public method, ~7 lines + Javadoc with usage example. Mirrors the body of scanBatches() minus the local stream allocation and the Data.importArrayStream step.
• No native code touched. The underlying JNI hook already existed; it was just not reachable from outside the class.
• Test testDatasetScannerExportArrowStream exercises the full path: caller allocates the C stream from its own RootAllocator, scanner fills the C struct, caller imports into an ArrowReader and validates batch contents (40 rows over 2 batches of 20).

Backwards compatibility

Pure addition. scanBatches(), schema(), countRows(), getStats(), close() all unchanged. No native ABI change.

Test plan

• ./mvnw test -Dtest=ScannerTest#testDatasetScannerExportArrowStream — passes locally (Java compile + spotless clean; full test run depends on a working lance-jni Rust build, which had an unrelated aws-smithy-types registry issue on my machine, so I'm relying on CI for the JNI-linked verification).
• Existing testDatasetScannerColumns covers the scanBatches() path so any regression in the shared openStream JNI call would surface there.

[sezruby]

@hamersaw @jackye1995 Could you review the PR? The PR is to support Lance Reader in Gluten/Spark. I'll open a lance-spark PR after this PR is merged. Thanks!