fix(index): preserve schema metadata when re-serializing loaded HNSW

[yanghua]

Summary

HNSW::to_batch() on a disk-loaded graph dropped all schema metadata
except the HNSW key, breaking the IVF partition-cache round-trip with
target schema is not superset of current schema.

When an HNSW partition is loaded from disk, its RecordBatch schema
inherits metadata keys from the index file (INDEX_METADATA_SCHEMA_KEY,
IVF_METADATA_KEY, …). The Loaded branch of to_batch() replaced the
schema metadata with a single-key map (HNSW_METADATA_KEY only), so when
RecordBatch::with_schema ran its superset check, the new schema was not
a superset of the original and the call errored out. This affects any
PartitionEntry<HNSW, *> (all quantizers) that re-serializes a loaded
index through the partition cache.

The fix clones the existing schema metadata and inserts/updates the HNSW
key, instead of building a fresh single-key map — guaranteeing the new
metadata is always a superset.

Changes

• rust/lance-index/src/vector/hnsw/builder.rs: in to_batch()'s
  Loaded branch, merge HNSW_METADATA_KEY into the retained batch's
  schema metadata rather than replacing it wholesale.
• Add test_to_batch_loaded_preserves_extra_schema_metadata: simulates
  the disk-load path by seeding extra schema metadata keys, then asserts
  to_batch() round-trips and preserves them.

[codecov]

Codecov Report

❌ Patch coverage is 97.61905% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
rust/lance-index/src/vector/hnsw/builder.rs	97.61%	0 Missing and 1 partial ⚠️

📢 Thoughts on this report? Let us know!