When someone needs to build a custom backend-specific GIS lookup, it
is much easier done if getting the spatial operator class happens in
a dedicated method (no need to rewrite the entire as_sql() method).

Without an explicit 'level', only messages at WARNING or higher
are handled. This makes the config consistent with the docs
which say, "The django catch-all logger sends all messages at
the INFO level or higher to the console."

Reverted "Fixes #24727 -- Prevented ClearableFileInput from masking
exceptions on Python 2" and added a regression test.

This reverts commit 5c412dd8.

Most likely this is a losing fight -- people seem to love this small
convention -- but at least the reasons for avoiding it will be
documented.

Refs #25356.

This commits lifts the restriction that the outermost atomic block must
be declared with savepoint=False. This restriction was overly cautious.

The logic that makes it safe not to create savepoints for inner blocks
also applies to the outermost block when autocommit is disabled and a
transaction is already active.

This makes it possible to use the ORM after set_autocommit(False).
Previously it didn't work because ORM write operations are protected
with atomic(savepoint=False).

Kept the docstring short because these objects aren't very well defined
and they're in the crosshairs of several refactorings.

The old names were downright confusing. Some seemed to mean the opposite
of what the class actually did.

The new names follow a consistent nomenclature:

    (Forward|Reverse)(ManyToOne|OneToOne|ManyToMany)Descriptor.

I mentioned combinations that do not exist in the docstring in order to
help people who would search for them in the code base.

Changed the poll / choices example to a more obvious parent / children.
I think that reduces the cognitive load.

At 2800 lines it was the largest module in the django package. This
commit brings it down to a more manageable 1620 lines.

Very small changes were performed to uniformize import style.

Too much field exclusions in form's construct_instance() in _post_clean()
could lead to some unexpected missing ForeignKey values.
Fixes a regression from 45e04993. Refs #13776.

The RegexURLResolver.resolve() method no longer returns a tuple.
It has returned a ResolverMatch object since commit e0fb90b2.

Thanks jwa and lamby for the suggestions, and timgraham and jarshwah
for their reviews.