fix: refresh embedded seekdb index after upsert

[webup]

Background

Embedded SeekDB 1.3+ builds vector indexes asynchronously. A caller that writes a record and immediately performs an ANN query can therefore miss that record until the pending index work is flushed.

langchain-oceanbase exposed this path because OceanbaseVectorStore writes through pyobvector.ObClient.upsert(). ObClient.insert() already flushes the embedded SeekDB index after a successful write, but upsert() did not.

Why Fix This in pyobvector

The missing behavior is in pyobvector's shared write primitive, not in LangChain. Aligning upsert() with insert() here:

• Gives every direct and indirect pyobvector consumer the same read-after-write ANN behavior.
• Keeps embedded-SeekDB details behind pyobvector's existing _flush_seekdb_index() hook.
• Avoids a LangChain-only workaround that would duplicate backend-specific logic, depend on a private pyobvector implementation detail, and leave other users exposed.

A future langchain-oceanbase dependency update can then add an end-to-end embedded HNSW regression test against the released pyobvector fix.

Change

Call the existing _flush_seekdb_index() hook after a successful ObClient.upsert() transaction, matching insert() behavior. The hook is already a no-op for non-embedded or unsupported backends, so remote OceanBase behavior is unchanged.

Validation

• Added a regression test that confirms upsert() invokes the embedded refresh hook.
• Verified the test fails before the change and passes afterward.
• Ran Python compilation and staged-diff whitespace checks.

The broader embedded suite was attempted but is not counted as validation: the local pylibseekdb 1.3.0.post2 environment fails during table creation before reaching upsert().