Bilateral speedup

[CattleProdigy]

At <dayjob> we noticed a pretty serious regression in performance in the bilateral filter when upgrading from 0.25 to 0.26. So I worked on speeding it up.

Not allocating more than necessary is by far the most significant improvement. The cartesian_product helper has some noticeable overhead (10s of ms). I also tried a well-known approximation for the exp(), I'm a little more skeptical of adding that one though. LMK

Baseline:
gray 500x500 r=5:
  median:   901.39ms  mean:   901.54ms  min:   900.65ms  max:   903.81ms
rgb  500x500 r=5:
  median:   939.59ms  mean:   940.08ms  min:   938.23ms  max:   942.73ms

Skip allocations and cartesian_product:
gray 500x500 r=5:
  median:   101.20ms  mean:   101.27ms  min:   101.12ms  max:   101.59ms
rgb  500x500 r=5:
  median:   155.27ms  mean:   155.30ms  min:   155.15ms  max:   155.53ms

exp approximation (+ previous optimizations)
gray 500x500 r=5:
  median:    74.99ms  mean:    75.01ms  min:    74.90ms  max:    75.14ms
rgb  500x500 r=5:
  median:   123.16ms  mean:   123.19ms  min:   122.96ms  max:   123.66ms

[cospectrum]

could you please add unit-tests (or better proptests) to check that the error of fast_exp is less than 4%?

[CattleProdigy]

could you please add unit-tests (or better proptests) to check that the error of fast_exp is less than 4%?

See changes here: c0bbe01

I tested all negative f32 values until -87 where f32::exp underflows anyway.