Unset params in CAGRA w/ build_algo = IVF-PQ not using heuristics

[ManasSingh12345]

Problem

CAGRA can tune IVF-PQ parameters (n_lists, pq_dim, n_probes, etc.) from dataset size and dimension — see ivf_pq_params and the Java copy in getCuVSIvfPqParams().

If a user sets only a few knobs (e.g. n_lists and pq_bits in a sweep), they expect the rest to be filled by those heuristics. That works with Strategy.HEURISTIC (≥ 5M vectors). It does not work with Strategy.CUSTOM, which is what you get when explicitly configuring IVF_PQ — unset fields fall back to fixed Builder defaults like n_lists=1024 and kmeans_trainset_fraction=0.5, regardless of dataset size.

Native code already does the right thing — heuristic baseline, then overlay only supplied fields (cagra.cpp) — but Java passes every field, clobbering the baseline. Only pq_dim=0 gets a partial escape hatch via calculate_pq_dim.

Expected

explicit value  >  dataset heuristic  >  static default

Per parameter. Match native behavior: baseline from ivf_pq_params(dataset), overlay only what the caller set.

Example

10M × 1536d — user sets only n_lists=5000, pq_bits=4:

Param	Should be	Today
kmeans_n_iters	10	20
kmeans_trainset_fraction	~0.1	0.5
n_probes	~8	20
search dtypes	fp16	fp32

(pq_dim accidentally works via 0 → calculate_pq_dim; other fields don't.)

Fix

1. Track explicitly set IVF-PQ keys.
2. Baseline = getCuVSIvfPqParams(N, D); overlay explicit keys only.