BUG: fix heuristic to switch from log scale to symlog scale

[neutrinoceros]

PR Summary

Fix #3791

Here's a simpler demo

import numpy as np
import yt

shape = (32, 16, 1)
a = np.linspace(0, 1, 16)
b = np.ones((32, 16))
c = np.reshape(a*b, shape)
data = {("gas", "density"): c}

ds = yt.load_uniform_grid(
    data,
    shape,
    bbox=np.array([[0.0, 5.0], [0, 1], [-0.1, +0.1]]),
)

p = yt.SlicePlot(ds, "z", "density")
p.save("/tmp/branch.png")

on main

this branch

While the original issue was only about 2D plots, I'm also fixing a similar bug in line plots that I found on the way

[neutrinoceros]

Not sure what's happening in the two tests that failed. If I can't figure it out quickly I'll just remove this from the milestone

[neutrinoceros]

Found that the failing tests were a consequence of #2517, and I discovered that the hack introduced back then isn't necessary after MPL 3.2 so I'm making it conditional. Now I'm expecting everything to stay green

[neutrinoceros]

I've revised my approach to intentionally change some reference images. We have a couple tests (the one that have been failing) for which we plot some "binary" noise (i.e. all values are 0 or 1). The current gold standard looks like this

which is not actually what we want: we see that all values plotted are equal (1), but we're masking zeros completely.
With the latest commit I'm getting this locally instead

Now I'm waiting to see if my changes accidentally affect any other cases, and wether it makes them better or worse.

[neutrinoceros]

This is now in the intented state. The only failing image tests are the ones I intend to update. I'll wait for an approval to bump them.

[neutrinoceros]

the diff here is that we convert to float even in the second case np.nanmin... specifically to strip out units, so we can easily increment with +1 later on.

[neutrinoceros]

The idea here is to intentionally overshoot zmax with the (excluded) upper boundary so that zmax can be used as a major tick location if it is itself an exact power of 10.

[matthewturk]

Makes sense.

[neutrinoceros]

necessary cleanup with matplotlib 3.5 where having a ticks that's above the actual maximal value creates a white region at the top end of the colorbar (see #3554)

[matthewturk]

Just so I understand, would it be possible to do this with np.logspace? I do not think we should in this PR, for continuity sake, however.

[neutrinoceros]

yes it should be possible to refactor this with np.logspace. I also agree that it's preferable to keep diffs minimal for a bugfix PR

[neutrinoceros]

Seems like the Jenkins job got killed because of network issues ?

hudson.remoting.ChannelClosedException: Channel "hudson.remoting.Channel@37a79109:JNLP4-connect connection from dxl1.ncsa.illinois.edu/141.142.193.70:39688": Remote call on JNLP4-connect connection from dxl1.ncsa.illinois.edu/141.142.193.70:39688 failed. The channel is closing down or has closed down

Let's try this again: @yt-fido test this please