bench: add qdrant comparison to vector benchmark

[kacy]

summary

• adds qdrant as an optional fourth comparison target in the vector similarity benchmark
• opt-in via --qdrant flag so existing usage is unchanged
• uses official qdrant/qdrant docker image and qdrant-client python SDK
• matches HNSW parameters (M=16, ef_construct=64, cosine) across all systems
• compatible with the ember-grpc benchmark added in main

what was tested

• cargo fmt and cargo clippy pass clean
• ran full 100k vector benchmark locally across ember, qdrant, and pgvector

local results (100k vectors, 128-dim, cosine, k=10)

metric	ember	pgvector	qdrant
insert (vectors/sec)	1,382	1,239	5,558
query (queries/sec)	1,776	519	337
query p99 (ms)	0.82	4.85	10.31
memory (MB)	29	193	31

ember query throughput is 5.3x faster than qdrant and 3.4x faster than pgvector at this scale. qdrant leads on insert throughput thanks to native batch upsert.

chromadb skipped — pydantic v1 is broken on python 3.14.

design considerations

• qdrant is opt-in (--qdrant flag) rather than always-on to keep docker optional for the default run
• --ember-only --qdrant works as a 2-system comparison (no chromadb/pgvector needed)
• uses query_points API (current qdrant SDK; search was removed)
• qdrant uses integer point IDs with string names in payload, since it doesn't support arbitrary string IDs