feat(redis): Add retrieve_online_documents_v2 support with Redis vect…

[falloficaruss]

What this PR does / why we need it:

Adds Redis 8 Vector Set support to the Redis online store for retrieve_online_documents_v2, so Feast can do ANN/vector similarity search without bypassing the SDK.

This PR:

• indexes vector fields into Redis Vector Sets using VADD
• queries vectors through VSIM
• supports COSINE and L2 metrics from Feast vector metadata
• keeps vector-set state aligned with hash writes and TTL behavior
• cleans up vector-set members when entities/tables are deleted
• adds unit coverage for indexing, deletion, and retrieval behavior

Which issue(s) this PR fixes:

Fixes #6461

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

[franciscojavierarceo]

Left a few inline comments on the Redis vector-set implementation. The main blockers are around the claimed L2 semantics and Redis < 8 capability detection.

[franciscojavierarceo]

This can report VADD as supported on Redis versions that do not have it. COMMAND INFO <name> returns a nil entry for unknown commands, so a response shaped like [None] still has len(info) > 0; writes would then proceed into VADD and fail instead of cleanly skipping/raising the intended unsupported-server path. Can we inspect the actual command entry/value, not just the container length?

[franciscojavierarceo]

This makes the advertised L2 mode behave like cosine/angular search. Normalizing indexed vectors and the query discards magnitude, so vectors like [2, 0] and [100, 0] become equivalent even though true L2 distances are very different. Since VSIM WITHSCORES returns Redis's similarity score, Redis should either reject L2 here or use a backend/query path that preserves real L2 semantics.

[falloficaruss]

If we prefer to try preserving true L2 semantics, we'd need to investigate whether Redis VSIM natively supports L2 and how to not normalize vectors in L2 mode, which is more complex. So, I will fix it by rejecting L2 here.

[franciscojavierarceo]

This cleanup is skipped when delete_table() returns early because no hash keys were found. That can leave the vector set behind after hash TTL expiry or partial cleanup. Could we delete vs_key before the if not all_keys: return path, or otherwise ensure table cleanup always removes the vector set?

[falloficaruss]

I've moved the vs_key deletion before the early-return guard so the vector set is always cleaned up when delete_table is called, regardless of whether hash keys are still present. DELETE on a non-existent key is a no-op, so it's safe when no vector set exists either