Instead of special-casing lognorm to only autoscale from positive
values, perform autoscaling from values that map to finite transformed
values. This ensures e.g. that make_norm_from_scale(LogitScale)
automatically does the right thing (autoscaling from values in [0, 1]).

This means that autoscale() and autoscale_None() are now slightly more
expensive (because the transform needs to be applied), so skip the call
to autoscale_None if not needed. However, note that these should
typically only be called once per norm anyways, so hopefully this isn't
a bottleneck.

(Another idea would be to usetrf.inverse().transform([-np.inf, np.inf])
as clipping bounds, but there are some tests usingx->x**2
/x->sqrt(x) as a test for FuncNorm, which 1. doesn't go all the way
to -inf, and 2. isn't even increasing for negative x's, so that idea
doesn't work.)

(See the first remaining issue of#20752, which is the motivation for this.)

PR Summary

PR Checklist

Tests and Styling

• Has pytest style unit tests (andpytest passes).
• IsFlake 8 compliant (installflake8-docstrings and runflake8 --docstring-convention=all).

Documentation

• New features are documented, with examples if plot related.
• New features have an entry indoc/users/next_whats_new/ (follow instructions in README.rst there).
• API changes documented indoc/api/next_api_changes/ (follow instructions in README.rst there).
• Documentation is sphinx and numpydoc compliant (the docs shouldbuild without error).