BUG: fix pixelize_cylinder

[neutrinoceros]

PR Summary

Fix #3781
I found that the actual problem was that the jumps in r and theta were constrained only by dx (horizontal pixel size) when they needed to be constained by dy too.

Using a variation of the script from the issue to show the change

import numpy as np
import yt

shape = (32, 64, 1)
a = np.linspace(0, 1, 64)
b = np.ones((32, 64))
c = np.reshape(a*b, shape)
data = {("gas", "density"): c}

theta_min = np.pi / 2 - 0.1

ds = yt.load_uniform_grid(
    data,
    shape,
    bbox=np.array([[0.5, 5.0], [theta_min, theta_min+0.1], [-0.1, +0.1]]),
    geometry="spherical",
)

yt.SlicePlot(ds, "phi", "density").save("/tmp/")

on the main branch

this branch

[neutrinoceros]

Switching to draft until I have time to go dig into failing tests.

[neutrinoceros]

Turns out the error is unrelated, and caused by a simple deprecation warning from SciPy 1.8
I'm pushing the fix here to make CI go green but I'll also make a separate PR to fix the problem for everyone
edit: this is #3783

[neutrinoceros]

In fact I'll leave this as a draft until I can test that my optimisations do not break other use cases

[neutrinoceros]

• failing tests on Jenkins may just need to be bumped (to be checked manually)
• I think my implementation can be optimised still: radial and angular increments in the loop are currently not dependent on theta, so they are de facto smaller than a pixel in most cases. This could be accelerated with a bit of geometry.

[matthewturk]

I think this is the correct fix. Thank you!

[neutrinoceros]

Got it back to a safe state with no overly eager optimisation after I found that I was reintroducing similar bugs. This should now be mergeable

[matthewturk]

It wasn't obvious to me that any optimizations were necessary.

[neutrinoceros]

Oh I just felt bad about making this routine a little slower since it's now doing more jumps. But I guess I shouldn't use a broken implementation as a baseline !