perf: chunked skip-scan in value_at_quantile (+63% read); supersedes #139

[fcostaoliveira]

Summary

Replaces the value_at_quantile prefix-sum scan with a chunked skip-scan. The old loop added one
count and branched on the running total for every element — a serial dependency chain. This sums a
fixed-size chunk (chunks_exact(8)) at a time — a reduction with no early exit that autovectorizes —
and skips the whole chunk while its subtotal cannot reach the target; only the single crossing chunk is
walked element-by-element. Counts are non-negative, so a chunk whose subtotal cannot reach the target
contains no crossing element: the result is identical to the linear scan for any input. No unsafe.

Supersedes and closes #139 — the chunked scan iterates chunks/elements rather than indexing, so it
already elides the per-element bounds check that #139 targeted; this is a strict superset of that win.

Benchmark

gnr1 (Intel Granite Rapids), single core, --release, same-session interleaved A/B (base main vs this branch), 2 rounds:

metric	base	this branch	Δ
value_at_percentile (Mq/s)	0.1739	0.2836	+62.9%
batch of 7 (7× singular, Mcalls basis)	24.8K	40.9K	+64.8%
record (M ops/s)	348.9	348.6	flat

Percentile results byte-identical across base and patch (sink/bsink unchanged).

Tests

• New white-box parity test comparing the chunked scan to a linear reference across all Counter
  widths (u8/u16/u32/u64) and a quantile sweep.
• A dedicated tail-crossing test: a sig=0 histogram (sub_bucket_half_count == 1) is the only geometry
  whose counts.len() is not a multiple of 8, so it forces a crossing element into the remainder()
  tail loop. (For sig >= 1, sub_bucket_half_count is a power of two ≥ 16, so counts.len() is always
  a multiple of 8 and the tail is unreachable.)
• cargo test, cargo fmt --check, cargo clippy all clean (zero new warnings); MSRV 1.64 / edition
  2018 unaffected (chunks_exact/remainder stabilized in 1.31); no new dependencies.