What this PR does / why we need it:

Three empty-input shapes reached internal exceptions, and the feature server turned each
into an HTTP 500:

entity_rows=[]              -> IndexError: list index out of range
entities={"driver_id": []}  -> KeyError: Missing join key values for keys: []
entities={}                 -> KeyError: 'pop from an empty set'

The middle case is a well-formed request: the join key was supplied, it simply holds
no values, which is what a caller sends when the upstream query matched nothing that run.

Three causes, fixed here:

• get_online_features built its columnar dict from entity_rows[0] on both the sync
  and async paths, so an empty list raised before anything else ran.
• _validate_entity_values ended with set_of_row_lengths.pop(); for a mapping with no
  columns the set is empty and pop() raised. It now reports zero rows.
• _get_unique_entities treated a join key present with zero values the same as one
  never supplied — even though the row-wise conversion immediately below it already
  returns empty results for that case (if not rowise: return (), (), 0). It now raises
  only when nothing at all was supplied for the view.

That last change is deliberately narrow, so existing behaviour is preserved: a partially
supplied key set still proceeds (covered by test_get_unique_entities_missing_join_key_success),
and a caller supplying nothing relevant still errors with the same message listing the
expected keys (covered by test_get_unique_entities_missing_all_join_keys_error).

The error is now MissingJoinKeyValuesException, carrying HTTP 400 so the server reports
a client error rather than a 500. It subclasses KeyError as well, since that is what
this condition raised before and callers catch it — so this is not a breaking change.

Resulting behaviour: join keys supplied but empty → 200 with an empty result and correct
feature-name metadata; nothing supplied → 400 naming the missing key.

Which issue(s) this PR fixes:

Fixes #6819

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

New sdk/python/tests/unit/online_store/test_empty_entity_rows.py. Four of its five
tests fail on master, reproducing all three symptoms above; the fifth asserts the new
exception is still a KeyError and carries 400.

Full sdk/python/tests/unit: 2612 passed, 24 skipped. The 6 failures and 28 errors
in that run are environmental and present on master too — PySpark worker/driver Python
version mismatch, a torch import, MongoDB testcontainers, and a missing CUDA library.

ruff check, ruff format --check and mypy are clean on all changed files.

Misc

An empty response was chosen over a 400 for the empty-but-present case, so batch-scoring
callers don't have to special-case "my filter matched nothing". It also lines up with
get_historical_features returning an empty frame for a zero-row entity_df. Happy to
switch to a 400 if maintainers prefer the stricter reading.