Rewrite the FFT peak search: block-parallel zoom refinement

At pad 40 the old path materialised a ~9 GB padded spectrum per row,
which collapsed the chunk planner to one worker and one rfft call with
workers=1 — synthesis ran single-threaded, ~1 hour per angle on real
files.

The padded spectrum is never materialised now. Each block of 512
waveforms gets a coarse rfft at next_fast_len(2*spf); every coarse bin
within 0.7 of its row's max (plus the DC-adjacent window, which coarse
DC suppression would otherwise blind) is refined onto the exact n_fft
grid by a small complex gemm. The selected bin is bit-identical to the
full padded argmax — enforced by test_zoom_identity, a 25-seed fuzz
test over adversarial spectra, and a clean golden-hash diff against the
pre-rewrite baseline across pads {1,2,4,8,40}, masked/unmasked, bg
on/off, int8/int16, and both backends.

Blocks fan out over a persistent thread pool; pyFFTW runs through
per-thread FFTW_MEASURE builder plans with wisdom persisted to
~/.cache/sras-viewer, and threadpoolctl clamps BLAS under the pool.
compute_rf_image(exact=True) (or SRAS_FFT_EXACT=1) keeps the reference
padded path for audits.

tools/bench_fft.py measures: pad 40, 16 cores, 8192x2500 synthetic —
exact serial 717 wf/s -> zoom pool 25100 wf/s (35x, pyFFTW backend;
19x scipy), every variant verified equal to the reference.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>

pyproject.toml

@@ -16,6 +16,8 @@ dependencies = [
     "scikit-image==0.26.0",
     # Faster rfft backend; the viewer falls back to scipy.fft without it.
     "pyFFTW==0.15.1",
+    # Clamps BLAS threading under the FFT worker pool.
+    "threadpoolctl==3.6.0",
 ]
 
 [project.optional-dependencies]