perf: prefetch counts[] ahead in the AVX2 percentile scan (follow-up to #138)

[fcostaoliveira]

Summary

Software-prefetch counts[] ahead of the AVX2 percentile scan
(get_value_from_idx_up_to_count_avx2) to hide L2/L3 load latency:

_mm_prefetch((const char*)&h->counts[idx + 64], _MM_HINT_T0);   /* 512 B / 4 iters ahead */

After the vector-accumulator widening, the scan is load-latency bound streaming the
~10s-of-KB counts[] array, so an explicit prefetch a few iterations ahead helps.

Stacked on #138. This branch contains #138's widening commit plus this one, so the diff
currently shows both. Please review/merge after #138 — once #138 lands, this PR reduces to
the single one-line prefetch commit. (I kept them as separate commits so each is reviewable on
its own.)

Benchmark

test/hdr_percentile_bench — hdr_value_at_percentile throughput, core-pinned, measured
base (this repo's AVX2 scan) vs +prefetch back-to-back in the same session, on two Intel
microarchitectures and both compilers:

µarch	Compiler	Base	+prefetch	Δ
Cascade Lake (Xeon Gold 6248)	gcc 11.4	0.37 M q/s	0.40 M q/s	+8%
Cascade Lake	clang 14.0	0.43 M q/s	0.43 M q/s	neutral
Granite Rapids	gcc 11.4	0.52 M q/s	0.56 M q/s	+7.7%
Granite Rapids	clang 14.0	0.53 M q/s	0.56 M q/s	+5.7%

gcc is consistently ~+8%; clang ranges neutral → +5.7% and never regresses. The write path
(hdr_histogram_perf) is a flat control on both µarchs (this change only touches the read scan).
The prefetch distance (64) is a reasonable default and could be tuned further per target.

Steps to reproduce

cmake -S . -B build -DCMAKE_BUILD_TYPE=RelWithDebInfo \
  -DHDR_HISTOGRAM_BUILD_PROGRAMS=ON -DHDR_HISTOGRAM_BUILD_BENCHMARK=ON
cmake --build build -j
taskset -c 8 ./build/test/hdr_percentile_bench     # base vs this branch: M queries/sec + sink

Correctness

• _mm_prefetch is a non-faulting hint — no bounds/UB implications, no change to the values read.
• ctest green (gcc and clang); the benchmark sink is byte-identical to base
  (17401860284404480), i.e. every percentile query returns exactly the same value.