What this PR does / why we need it:

Every other online store returns a feature value as it was written, whatever its type, because
Redis, Postgres, SQLite, DynamoDB, Elasticsearch, Qdrant and the rest store the serialized
ValueProto. The Milvus online store keeps one column per feature, and with the default
configuration it could not return most non-numeric types. On upstream master with Milvus Lite:

The write path now stores every scalar natively or as text, bytes as base64, and every other
value as the base64 of its ValueProto, for features and entity columns alike, and one helper
rebuilds the ValueProto for reads and searches from every declared field type. UnixTimestamp gets a column. Only a field with vector_index=True becomes a vector,
with the dimension from Field.vector_length before embedding_dim; the Elasticsearch, Postgres,
Qdrant and MongoDB stores already pick the vector field this way. An empty text in a non-string
column reads as null. Array(Bool) with vector_index=True now raises at collection creation:
it was mapped to a binary vector but written unpacked and indexed with a float metric, so it could
not work.

A missing value is still written as the column's default when the column cannot hold NULL. The
new milvus.nullable: "true" field tag makes a column nullable, so a missing value is stored and
read as null, as every other store does:

Field(name="rating", dtype=Float32, tags={"milvus.nullable": "true"})

Nothing here needs a flag. Every change fixes a call that failed or returned a wrong value, the
read fixes apply to existing collections at once, and the new columns and the tag only apply when
Feast creates a collection, like the partition key.

Which issue(s) this PR fixes:

N/A

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

• Unit (Milvus Lite): thirteen Feast values, empty bytes included, round-trip through
  online_write_batch and online_read, with and without native numeric types; a missing value
  in a text column reads as null; a milvus.nullable column stores null and a plain column keeps
  its default; search hits decode every type the same way as reads; Bytes and UnixTimestamp
  entity columns round-trip.

• Unit (mocked client): a numeric array without vector_index is a VARCHAR column, a
  vector_index field takes its vector_length, UnixTimestamp gets a column; the tag is
  rejected on entity, vector and partition key fields and for an invalid value; Array(Bool) with
  vector_index is rejected.

• Server: the same round-trip and the nullable column against a real server, with and without
  native numeric types.

Unit tests run on Milvus Lite 3.2.1 (pymilvus 3.0.2). The server tests in
sdk/python/tests/integration/online_store/test_milvus_remote.py pass against a local Milvus
3.0.2 server. The existing Milvus unit and server tests pass unchanged.

Misc

Follows #7082, #7083 and #7084 and Simon Hearne's series #6882, #6895, #6917, #6921 and #6940.
The next PR builds on this one: native ARRAY and JSON columns and scalar index tags behind
enable_openai_compatible_store.