What this PR does / why we need it:

The SQLite online store answers retrieve_online_documents and retrieve_online_documents_v2 (and so the feature server's /search) by copying the feature view's rows into a vec0 (vector) or FTS5 (keyword) table and querying that. The table lives in the online store database and is never cleared, so every search adds another copy of every row:

• Vector search (vector_enabled: true) works only once. The next retrieve_online_documents_v2 call fails with sqlite3.OperationalError: UNIQUE constraint failed on vec_table primary key, and retrieve_online_documents with table vec_table already exists. Once a write commits the leftover rows, even the first search of a new process fails.
• Keyword search (text_search_enabled: true) returns fewer documents on every call, because copies of the best matches fill top_k, and it keeps matching text that has since been overwritten.
• The search also leaves its transaction open, so writes from any other connection (e.g. a feast materialize next to feast serve) fail with database is locked until the searching process writes something itself.

Repro on master, same data and query each time:

keyword search "sentence", top_k=3
  call 1: Title 0, Title 1, Title 2
  call 2: Title 0, Title 1
  call 3: Title 0
  after item 0 is rewritten without "sentence": still returns Title 0

vector search, top_k=3 (Python 3.10, sqlite-vec)
  call 1: doc 0, doc 1, doc 2
  call 2: OperationalError: UNIQUE constraint failed on vec_table primary key

This PR builds the index in the connection's temp schema, drops and refills it for every search, and commits it right away. Every call now returns the same top_k from the current rows, a rewritten document no longer matches its old text, and nothing is left in the online store file or locked. The vector index build is shared by both APIs.

Which issue(s) this PR fixes:

No existing issue.

Checks

• I've made sure the tests are passing.
• My commits are signed off (git commit -s)
• My PR title follows conventional commits format

Testing Strategy

• Unit tests
• Integration tests
• Manual tests
• Testing is not required for this change

Misc

Added two tests to sdk/python/tests/unit/online_store/test_online_retrieval.py; both fail on master and pass with this change:

• test_sqlite_get_online_documents_v2_search_is_repeatable repeats a keyword search and rewrites a document.
• test_sqlite_get_online_documents_is_repeatable runs retrieve_online_documents_v2 and retrieve_online_documents twice each (Python 3.10 only, like the existing sqlite-vec test).

All sqlite tests in that file pass on Python 3.10 with SQLite 3.53 (vec0 KNN queries need SQLite >= 3.41 for LIMIT, as on ubuntu-latest), as does test_sqlite_get_online_documents when run directly. I did not run the Milvus tests (milvus-lite has no Windows build). ruff and mypy on sqlite.py are clean.