Closes #6179. Part of #2728.

What this does

Routes every Qdrant collection reference through one helper, so that with
registry.enable_online_feature_view_versioning: true each feature view version gets
its own collection (driver_stats_v2) instead of all versions sharing one. Applied to
the write, create, update, teardown and document-retrieval paths.

One deliberate deviation from the FAISS/Milvus precedent

Those stores use compute_table_id(project, table, versioning), which yields
{project}_{name}[_v{N}]. This PR uses compute_versioned_name instead, which yields
{name}[_v{N}] with no project prefix.

The reason is that Qdrant collections have always been named by the bare table.name —
unlike Milvus, which already had the {project}_ prefix before versioning was added to it,
so gaining the flag there was non-breaking. Switching Qdrant to compute_table_id would
rename every collection in every existing deployment and orphan their data. With versioning
disabled the collection name here is byte-identical to today's, which
test_unversioned_store_still_round_trips pins.

Happy to switch to the project-prefixed form if you'd rather have consistency across
stores and want to handle the migration — just say so.

Why online_read is not added to the versioned-read allowlist

I did not add QdrantOnlineStore to OnlineStore._is_versioned_read_supported(), because
QdrantOnlineStore.online_read cannot currently serve any read, versioned or not. It looks
like it has never been exercised — there is no unit test for it, and it fails on two
independent counts before reaching Qdrant:

1. it passes raw EntityKeyProto objects into models.MatchAny(any=entity_keys), which
   pydantic rejects (Input should be a valid string), and
2. it reads collection_name=config.online_store.collection_name, but neither
   QdrantOnlineStoreConfig nor VectorStoreConfig defines collection_name, so that
   line raises AttributeError.

It also returns one entry per stored point holding a base64 str, where the
OnlineStore.online_read contract (see sqlite.py) is one entry per requested entity key,
in order, holding ValueProto values, with (None, None) for a miss.

Reproduction, stock main, no external service — an in-process Qdrant is enough:

from datetime import datetime, timedelta
from feast import Entity, FeatureView, Field, FileSource, RepoConfig
from feast.types import Int64, String
from feast.value_type import ValueType
from feast.protos.feast.types.EntityKey_pb2 import EntityKey as EntityKeyProto
from feast.protos.feast.types.Value_pb2 import Value as ValueProto
from feast.infra.online_stores.qdrant_online_store.qdrant import (
    QdrantOnlineStore, QdrantOnlineStoreConfig,
)

cfg = RepoConfig(
    project="proj",
    online_store=QdrantOnlineStoreConfig(
        type="qdrant", location=":memory:", vector_enabled=True, similarity="cosine"
    ),
    registry="dummy",
    entity_key_serialization_version=3,
)
fv = FeatureView(
    name="stats",
    entities=[Entity(name="user_id", value_type=ValueType.INT64)],
    ttl=timedelta(days=1),
    schema=[Field(name="user_id", dtype=Int64), Field(name="f1", dtype=String)],
    source=FileSource(path="t.parquet", timestamp_field="event_timestamp"),
)
store = QdrantOnlineStore()
store.update(cfg, [], [fv], [], [], partial=False)

ek = EntityKeyProto(join_keys=["user_id"], entity_values=[ValueProto(int64_val=1)])
store.online_write_batch(
    cfg, fv, [(ek, {"f1": ValueProto(string_val="v1")}, datetime(2024, 1, 1), None)], None
)          # -> succeeds

store.online_read(cfg, fv, [ek], ["f1"])
# pydantic_core._pydantic_core.ValidationError: 2 validation errors for MatchAny
#   any.list[str].0  Input should be a valid string ... input_type=EntityKey

Fixing that properly means deciding how entity_key should be stored — it is currently
written into the payload as raw bytes from serialize_entity_key, which MatchAny cannot
filter on, and which retrieve_online_documents then reads back through
str(payload.get("entity_key")), producing a bytes repr rather than the value
_build_retrieve_online_document_record expects. That is a storage-format call I did not
want to make unilaterally inside a versioning PR, so I have kept it out of scope. Happy to
open a separate issue, or to take it in a follow-up once you've said which encoding you want.

So this PR delivers the versioned collection namespace; versioned scalar reads stay
correctly gated behind VersionedOnlineReadNotSupported until online_read works.

Tests

New sdk/python/tests/unit/infra/online_store/test_qdrant_online_store.py, 6 tests, running
against an in-process Qdrant (location=":memory:") — no service required, and skipped via
importorskip when qdrant-client is absent.

• naming: unversioned unchanged, version tag ignored when the flag is off, _v2 when on
• test_versions_write_to_separate_collections — v1 and v2 collections both created, one
  point each, no cross-contamination
• test_teardown_removes_only_the_targeted_version
• test_unversioned_store_still_round_trips — the control

Verified red-before/green-after: with the change reverted, the two versioning behaviour
tests fail while the unversioned control still passes, so they are testing the change rather
than the setup.

Regression check across the 56 unit test files touching online stores: 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 both
files.

🤖 Written with Claude Code (Claude Opus 5), reviewed by @arose26.