GeezerVison

Closes #6174. Part of #2728.

What this does

Threads registry.enable_online_feature_view_versioning through _map_name, which now
delegates to compute_table_id, so that with versioning enabled each feature view version
gets its own Hazelcast map (test_project_driver_stats_v2) instead of all versions sharing
one. Applied to every call site — online_write_batch, online_read, update (both the
keep and delete loops) and teardown.

Hazelcast already namespaced as {project}_{table.name}, so this is the same
compute_table_id convention the merged Milvus (#6330) and FAISS (#6256) stores use — no
deviation. With versioning disabled the map name is byte-identical to today's, pinned by the
versioning=False cases.

One small refactor, and why it is necessary

update() and teardown() previously called _map_name(...) inline inside f-strings:

client.sql.execute(f"DELETE FROM {_map_name(config.project, table)}").result()

With the extra argument that call no longer fits on one line, and line breaks inside an
f-string replacement field are a Python 3.12 feature (PEP 701) while this package declares
python_requires=">=3.10". So the resolved name is bound to a local first and the f-strings
just interpolate it. ruff with the project's py310 target rejects the inline form, which
is how I caught it.

Versioned-read allowlist: added, after checking

HazelcastOnlineStore is added to OnlineStore._is_versioned_read_supported(). I read the
path rather than assuming: online_read builds entity_keys_str from the requested keys in
order, then iterates it appending (event_ts, entry) for a hit and (None, None) for a miss —
so it returns one aligned entry per requested key, which is the contract sqlite.py implements
and the property versioned reads depend on. test_read_returns_one_entry_per_requested_key
pins it directly rather than taking my word for it.

Write and read also both serialize entity keys at version 3, so they agree on keys and the
change does not desynchronise them.

Drive-by: a stale error message

VersionedOnlineReadNotSupported listed "SQLite, PostgreSQL, MySQL, FAISS, Redis, and
DynamoDB", but _is_versioned_read_supported() has included MilvusOnlineStore since #6330,
which updated online_store.py without touching errors.py. So the message has been telling
Milvus users their store is unsupported while the code supported it. Added Milvus alongside
Hazelcast — happy to split that into its own PR if you'd prefer.

Tests

New sdk/python/tests/unit/infra/online_store/test_hazelcast_versioning.py, 15 tests,
MagicMock-based in the same style as the merged test_milvus_versioning.py, and
importorskip-guarded on hazelcast.

• naming: unversioned unchanged; version ignored while the flag is off; _v2 when on;
  projection.version_tag takes priority over current_version_number; version 0 gets no
  suffix; two versions never collide
• routing: write, read, update (create and drop) and teardown all operate on the
  versioned map — write is parametrized so the unversioned case is asserted as the control
• allowlist: _is_versioned_read_supported() is true, _check_versioned_read_support no
  longer raises, and the alignment property the allowlist rests on is asserted

Verified red-before/green-after: with the three source files reverted, seven of these fail,
while the versioning=False control and the read-alignment test still pass — the latter
correctly, since alignment is a pre-existing property this PR relies on rather than introduces.

Regression check across the 55 unit test files touching online stores or versioning: the set of
failing and erroring tests is identical with and without this change (diff of the sorted
FAILED/ERROR lines is empty). Those pre-existing failures are missing optional dependencies
in my environment. ruff check, ruff format and mypy are clean on all four files.

🤖 Written with Claude Code (Claude Opus 5), reviewed by @Arose26.