When blob columns flow through Spark's shuffle (e.g., INSERT INTO
target SELECT ... FROM source_a JOIN source_b), the actual blob data
was previously lost. This PR introduces a blob reference mechanism
that preserves blob data through shuffle without materializing the
full blob bytes.

Read side: blob columns serialize compact ~100-byte BlobReference
descriptors (LANCEREF magic + dataset URI + column name + row address)
instead of empty bytes. The scanner requests _rowaddr when blob columns
are present and strips it from the output.

Write side: LargeBinaryWriter detects BlobReference headers, buffers
them during setValue(), then batch-resolves all references in finish()
via a single takeBlobs() call per (dataset, column) group. Dataset
instances are cached across batches for the task lifetime.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Addresses review feedback on the blob JOIN/INSERT preservation path and
makes vended-credential auto-refresh work for blob sources.

Correctness:
- LargeBinaryWriter now buffers all per-row values and emits the vector in
  a single ascending pass at finish(), instead of back-filling resolved
  blobs out of order (which corrupts a LargeVarBinaryVector's offset
  buffer). resolveBatch() returns an index->bytes map rather than writing
  into the vector.

Resource lifecycle:
- The BlobReferenceResolver is now created once per write task, shared
  across batches/fragments, and closed at LanceDataWriter teardown —
  fixing a leak of native source datasets (one per blob column per batch).

Credentials (the main design change):
- New LanceBlobSourceContextRule optimizer rule collects each blob source
  table's BlobSourceContext (read options + namespace config) on the
  driver and stashes them, keyed by source URI, in the write command's
  options. LanceDataset.newWriteBuilder decodes them and threads them to
  the per-task resolver, which reopens sources via
  Utils.openDatasetBuilder().runtimeNamespace(...) so vended credentials
  keep auto-refreshing — mirroring compaction/index. No global registry,
  no per-row shuffle bloat. Registered in LanceSparkSessionExtensions.
- BlobReferenceResolver no longer opens datasets directly; falls back to
  open-by-URI when no context is present (local sources / extension off).

Performance:
- BlobStructAccessor precomputes the constant reference prefix once per
  batch in setBlobReferenceContext; the per-row path only appends the
  8-byte rowAddress.

Misc:
- Extract LargeBinaryWriter to its own file.
- Fold the new constructor arguments into single canonical constructors
  rather than parallel overloads; update affected tests.
- BaseBlobJoinTest enables the SQL extension so the rule path is covered.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Address PR review feedback on the blob join-preservation path:

- BlobReferenceResolver.resolveBatch: deduplicate row addresses within
  each (datasetUri, columnName) group and fan resolved bytes back out by
  address instead of by list position. Guard against takeBlobs count
  mismatch and null elements by failing loudly rather than writing wrong
  bytes. Document the takeBlobs ordering/1:1 contract.
- BlobStructAccessor.getBlobReference: test blob size with the primitive
  UInt8Vector accessor instead of boxing through a per-row BigInteger on
  the scan hot path.
- LargeBinaryWriter: buffer lazily — write rows directly until the first
  blob reference, then buffer only the tail. The common non-blob binary
  case now buffers nothing.

Tests (written as JUnit 5 in Scala so surefire actually executes them;
the existing ScalaTest *Suite classes are not picked up by the build):
- LargeBinaryWriterTest: direct/buffered ordering, reference resolution
  at correct indices, IOException->RuntimeException, reset.
- LanceBlobSourceContextRuleTest: non-Lance target no-op, blob-free
  source no-op, positive annotation, existing-key guard, idempotence.
- BaseBlobJoinTest: one-to-many JOIN test exercising resolver dedup.

Verified across Spark 3.5 and 4.1 (Scala 2.13).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>