feat(index): add IVF+PQ compression flag for the FAISS build

[dex0shubham]

Closes #153.

Motivation

The FAISS backend uses IndexIVFFlat, storing every vector at full float32
precision — ~245 GB of RAM for a 30 M-vector corpus at dim 2048. That's out of
reach for air-gapped or Docker-free deployments that can't run a Qdrant sidecar,
even though Qdrant users can already compress via --qdrant-quantization-config.
This adds the equivalent knob on the FAISS side.

Change

Opt-in --pq-m / --pq-nbits flags select IndexIVFPQ (Product Quantization)
instead of IndexIVFFlat. At pq_m=64, nbits=8 each vector shrinks from 8192
to 64 bytes (~128x, ~1.9 GB for 30 M vectors) for a typical small recall cost.

• embed/src/pixelrag_embed/index.py — build_ivf takes pq_m (default 0 =
  unchanged IVFFlat) and pq_nbits (4 or 8). When pq_m > 0 it builds
  IndexIVFPQ; the existing .train() call trains the PQ codebooks too, so no
  new training path. summary.json records index_type / pq_m / pq_nbits
  so the served index is self-describing.
• index/src/pixelrag_index/pipelines.py — forwards index.pq_m /
  index.pq_nbits from pixelrag.yaml to the build command.

index:
  backend: faiss
  pq_m: 64      # sub-quantizers; must divide the embedding dim evenly
  pq_nbits: 8   # 4 or 8 (default 8)

One fix vs. the proposal

The issue's sketch was faiss.IndexIVFPQ(quantizer, dim, nlist, pq_m, pq_nbits),
which defaults to METRIC_L2 — wrong for the IP-normalized embeddings this
index uses, and a silent ranking regression. This PR passes metric_type
explicitly and a test locks it in.

Validation: pq_m must divide the embedding dim evenly, else a clear
ValueError (FAISS would otherwise assert deep in C++).

Tests

tests/test_ivfpq_index.py (plain pytest, faiss via importorskip, no model
— builds a real index from a synthetic shard):

• IVFPQ index is built, trained, has ntotal == n, is searchable, and its
  metric_type is inner product (guards the L2-default bug).
• summary.json records index_type="ivfpq" and the PQ params.
• pq_m not dividing the dim raises ValueError.
• pq_m=0 (default) still produces an IndexIVFFlat — no behavior change.

All pass; ruff check / ruff format --check clean.

[vercel]

@dex0shubham is attempting to deploy a commit to the andylizf's projects Team on Vercel.

A member of the Team first needs to authorize it.