PERF: optimize cylindrical pixelizer routine

[neutrinoceros]

PR Summary

Yet another alternative to #5018 and #5034
This separates filtering of pixels that only partially intersect with the domain and aims to do a minimal amount of work there. On my system I'm getting similar performances as compared to #5034, but the goal here is to not change the observable behavior. Let's see how this does in CI 🤞🏻

[neutrinoceros]

I realized that there's still a defect with my approach: I'm only shaving the pixilization from the outside in, missing most (if not all) pixels on the inside of the radial domain. I think I can figure this out too but it'll take more time than I have left today, so it'll wait for another day. I also don't think we have any image test that would reveal this bug as we use rmin=0 everywhere. I think we should start by adding a test case for this.

I also see a segfault on windows, which is probably not actually platform specific, and needs to be addressed in any case.

[neutrinoceros]

I fixed the segfault, which was actually trivial to locate by re-enabling bound checking for debugging.

[neutrinoceros]

Adding the missing tests in #5036

[neutrinoceros]

rebased with #5036 merged. I expect one of the two new image tests to fail.

[neutrinoceros]

Hum, I haven't been able to figure out a trick to fix "pixel shaving" on the inner side without compromising performance yet, so I'm starting to think it'd be preferable to go with #5018 first and revisit this after the 4.4.0 release. Does this suite you, @chrishavlin ?

[chrishavlin]

Ya, that's fine by me! We can also hold #5018 to after 4.4.0 if that would make the effort in this PR easier.