Implement Kapur's algorithm for binary thresholding (#696)

[o-tho]

Kapur's algorithm uses an entropy approach to find a good segmentation threshold for gray images.

This implementation of Kapur's algorithm sticks closely to the notation of the original article (https://doi.org/10.1016/0734-189X(85)90125-2), so it is much easier to check for correctness. We also avoid a double loop, so this only needs 4*256 computations instead of 256^2.

[cospectrum]

Why not replace f64::EPSILON with 0.0?

[o-tho]

That's a very good question. While playing around with different possible implementations I got concerned about summing up floats of different orders of magnitude and wanted to give myself some safety margin, but the p*ln(p) are completely stable and I never changed it back.

[theotherphil]

Thanks! Please can you make the change suggested by clippy (not obviously an improvement, but not worse and this is easier than ad-hoc disabling of lints) and then I’ll merge.

[cospectrum]

I will make a PR to lock the nightly channel for clippy

[o-tho]

Thanks! Please can you make the change suggested by clippy (not obviously an improvement, but not worse and this is easier than ad-hoc disabling of lints) and then I’ll merge.

I can absolutely do that, but just for the record: The clippy change is about code that has nothing to do with the PR. You still want me to push it here?

[theotherphil]

Ah, I hadn’t spotted that. Fixed by @cospectrum in #700.

Thanks for the new feature.