viz: Should a linear scaling be used automatically when the value range is narrow ?

[neutrinoceros]

Trying to fix #3229 makes me think that it could be an opportunity to make norm selection a little more clear, and in particular it seems like a good time to propose a new rule.
So here's an example to illustrate how our current scale selection rules can be damaging for data readability

import numpy as np
import yt

d = {'my_fake_density': np.random.random_integers(1, 3, size=(64,64,64))}

ds = yt.load_uniform_grid(d, [64,64,64])
p = yt.SlicePlot(ds, "z", ('stream', 'my_fake_density'))
p.set_log(('stream', 'my_fake_density'), True, linthresh=1.e-3)
p.set_zlim(('stream', 'my_fake_density'), -1, 1)
p.save()

output:

The default colorscale is actually pretty obscure because a log norm isn't appropriate here.
(note: I've used my branch #3536 to produce this picture to avoid distraction from an unrelated colorbar bug)

Using yt's main branch

)

As an easy fix to the issue above, I'm proposing that we use linear scaling by default instead of log on the condition that the value range is below 2 orders of magnitude. I believe this can be achieved at the same time as a fix to #3229.

Of course this criterion isn't sufficient and it would also make sense to compare the variance of the image VS the variance of its log, but I don't really like the idea of piling up layers of magic here, as the existing norm-selection mechanism is already proving difficult to maintain against a large range of matplotlib versions.
In an ideal world, we should expose the color norm completely so all of matplotlib's flexibility is available to our users.
I would like to do that in the future, but for now I want to focus on having sensible defaults.

[neutrinoceros]

arguably the example in the OP is terribly malformed, I don't think it's illustrating the point correctly now so I'd rather close it and maybe reopen when things stabilise with matplotlib 3.5