[BUG] bitset-filter result- padding quirk in cuvs::neighbors::ivf_pq::search

Describe the bug
When the number of rows passing the bitset filter is strictly less than the
requested top-k, IVF-PQ does not write the (-1, +inf) sentinel into the
remaining result slots. Instead it pads them with filter-EXCLUDED nearest
neighbors (rows whose bit in the filter bitset is 0). The caller cannot tell
these padding entries apart from real, filter-passing matches.
IVF-Flat and CAGRA on the same input correctly emit the sentinel for the
excluded slots — so this looks specific to the IVF-PQ search kernel's
handling of popcount(filter) < k.

Observation: in our run the distance slot is already FLT_MAX (+inf) for
the leaked entries — cuVS evidently knows those slots are empty and writes
the sentinel distance, but it forgets to also overwrite the matching
neighbor index with -1. So the fix on the cuVS side may be a single store
in the same kernel codepath that already writes the sentinel distance.

Steps/Code to reproduce bug
Follow this guide http://matthewrocklin.com/blog/work/2018/02/28/minimal-bug-reports to craft a minimal bug report. This helps us reproduce the issue you're having and resolve the issue more quickly.

reproducer program attached.

Expected behavior
A clear and concise description of what you expected to happen.

invalid results should be filtered out.

Environment details (please complete the following information):

• Environment location: [Bare-metal, Docker, Cloud(specify cloud provider)]
• Method of RAFT install: [conda, Docker, or from source]
  • If method of install is [Docker], provide docker pull & docker run commands used

Reproduces on libcuvs / libraft 26.02.000 + CUDA 13.0 against an
RTX 5070 Laptop GPU.

Additional context
Add any other context about the problem here.

./test_ivfpq_filter
==============================================================
cuVS IVF-PQ bitset-filter padding reproducer
  rows=200 dim=8 n_lists=4 pq_dim=4 n_probes=4 top_k=5
  filter popcount = 2  (passing IDs: 0 2)
  query = (1,1,...,1).  Distance from query to each row:
    row 0 (passing): 0.0   row 1 (excluded): 8.0   row 2 (passing): 32.0
    row 3+ (excluded): ~10000+
--------------------------------------------------------------
Expected (correct):     neighbors = [   0,    2,   -1,   -1,   -1]
--------------------------------------------------------------
  raw cuVS:              neighbors = [   0,    2,    1,   40,   41]   distances = [   0.0,   32.0,    inf,    inf,    inf]
  ==> 3 filter-excluded id(s) leaked into the result.  <-- bug
  after host post-filter: neighbors = [   0,    2,   -1,   -1,   -1]   distances = [   0.0,   32.0,   +inf,   +inf,   +inf]
--------------------------------------------------------------
RESULT: bug reproduced; host post-filter recovers the expected sentinel-padded result.

Start with the attached test_ivfpq_filter reproducer and trace the cuVS IVF-PQ search kernel path used when popcount(filter) is less than top-k. Compare its handling with IVF-Flat and CAGRA, then verify that excluded result slots contain neighbor index -1 alongside the existing +inf distance in the reproducer output.