Wrap the raw HTTP-backed AsyncBuffer rather than the cached layer so the
stats bar reports actual network fetches (not cache hits). Adds two new
metrics that surface the per-query bottleneck:

  net wait:       summed time across all in-flight slice() calls
  max concurrent: peak in-flight requests (confirms h3 parallelism)

The previous counter wrapped the cached buffer, so cache hits were
counted as fetches and the numbers were inflated 2x.

Bump hypvector dep to 0.1.1 and call prefetchBinary at page load so
phase-1 Hamming scans run from memory. Cuts ~30 fetches per query and
significantly tightens p90 over the CloudFront/WAN cold path.

The 7.5 MB binary column download happens once during the 'loading
vector index...' status, then every subsequent query skips its
binary parquet fetches entirely.