Depends on LadybugDB/extensions#28

Some notes that may be helpful to people reviewing this code:

Vector extension (quantized HNSW)

Before this commit, at submodule 1bf007bca:

• Quantized embeddings were NOT persisted to a dedicated disk table. QuantizedOnDiskEmbeddings read raw float embeddings from the node table and quantized them in memory on every access. There was no quantizedEmbeddingsTableID stored in HNSWStorageInfo.
• The cache was a simple member variable (quantizedEmbeddingsCache / quantizedEmbeddingsCacheNumNodes) that was bluntly reset after every commitInsert() or finalize() call.
• Checkpoint did not touch quantized data — there was no quantized embeddings table to checkpoint.
• No concurrency handling — the cache was not shared across threads or queries in a safe way.

The new code at c0b7d47 adds:

• Persistent quantized embeddings table: Quantized embeddings are now stored to a dedicated NodeTable (quantizedEmbeddingsTable), serialized in HNSWStorageInfo, and checkpointed. - Thread-safe shared cache: The QuantizedEmbeddingsCacheState uses a mutex and reference-counted shared pointer, allowing multiple concurrent queries to share the same cached column.
• Transaction-aware invalidation: Cache is only invalidated on actual commit, not on every write.
• Version-gated reads: The cache carries a version (commit timestamp) allowing readers to determine if the cache is up-to-date for their snapshot.