Fix kvrocks: pipeline smembers to eliminate N+1 Redis round trips

[t0kubetsu]

Summary

Follow-up to the closed PR #155. The not/nt modifier change has been dropped per maintainer feedback — that is intentional and will be addressed in a future parenthesis/syntax refactor.

This PR contains only the N+1 fix, which is independent of the modifier work.

Problem

In get_uids_by_criteria, get_uids_by_criteria_scoped, and _get_uids_for_http_headval, the scan_iter loops issued one synchronous self.r.smembers(key) call per matching key. On a large index, a field.like:value query could produce thousands of serial Redis round trips in a single search request.

Change

Two-phase approach in all three methods:

1. Collect all matching keys from scan_iter (key filtering only, no data reads)
2. Issue all smembers calls in a single pipeline(transaction=False) batch

# Before — N+1 round trips
for key in self.r.scan_iter(f"{base_field}:*"):
    if matches(key):
        matching_uids.update(self.r.smembers(key).intersection(partial_result))

# After — 1 pipeline per query
matching_keys = [k for k in self.r.scan_iter(f"{base_field}:*", count=1000) if matches(k)]
if matching_keys:
    pipe = self.r.pipeline(transaction=False)
    for k in matching_keys: pipe.smembers(k)
    for members in pipe.execute():
        matching_uids.update(set(members).intersection(partial_result))

Also adds the missing count=1000 hint to the two scan_iter calls that were omitted.

Test plan

• A field.like:value query returns the same results as before
• A field.begin:value query returns the same results as before
• Redis monitor shows a single PIPELINE batch instead of N individual SMEMBERS commands