Tone match ran about 3,700 small torch operations per picture. PyTorch
spreads each one over every CPU it sees, and a cloud container can show it
far more CPUs than it may use, so each one waited its turn: on a RunPod
container it added 37 to 146 seconds to a 1.6 MP picture.
The blurs now run once over every layer they need, and the fit, the
held-out check and the curve inversion run over whole arrays: the band's
samples are summed per knot segment (sorted once, one reduction per pass
for all three fits), every curve is solved in one batched call, and the
curves are read and inverted over the whole picture at once. That leaves
about 18 multi-threaded operations. A single picture reuses the values it
found while measuring when the correction is applied.
The result matches the old code: same decisions, and at most 0.0002 of one
8-bit level apart on real runs (the fit's sums are only added in another
order). The old code is kept in the tests as a reference, and a new test
pins the two together on turned, straight, dark, thin-feather, clip and
alpha cases.
