Thanks to Keryn Knight, Curtis Maloney and Tim Graham for their reviews.

Refs #24704.

With this change, it's expected to survive anything except errors
that make it impossible to import the settings. It's too complex
to fallback to a sensible behavior with a broken settings module.

Harcoding things about runserver in ManagementUtility.execute is
atrocious but it's the only way out of the chicken'n'egg problem:
the current implementation of the autoreloader primarily watches
imported Python modules -- and then a few other things that were
bolted on top of this design -- but we want it to kick in even if
the project contains import-time errors and django.setup() fails.

At some point we should throw away this code and replace it by an
off-the-shelf autoreloader that watches the working directory and
re-runs `django-admin runserver` whenever something changes.

This also fixes a test failure on Python 2 when Django is installed in a
non-ASCII path. This problem cannot happen on Python 3.

* Added helpers to test uncached and cached access.
* Fixed test_project_root_locale: it duplicated test_locale_paths_setting.
* Rewrote test_only_new_files: test more cases.

* When some old files contain errors, the second call to
  gen_filenames() should return them.
* When some new files contain errors, the first call to
  gen_filenames(only_new=True) should return them.

Changed the way makemessages invokes xgettext from one call per
translatable file to one call per locale directory (using --files-from).
This allows to avoid https://savannah.gnu.org/bugs/index.php?35027 and,
as a positive side effect, speeds up localization build.

Previously, messages of ERROR level or higher were printed to
the console.

This speeds up import of 'django.core.validators' which can save a
few hundred milliseconds when importing the module for the first
time. It can be a significant speedup to the django-admin command.

Fixed #25299 -- Fixed crash with ModelAdmin.list_display value that clashes with a model reverse accessor.

Fixed a regression introduced in e1427cc6 when running tests in reverse order.

Introduced an AbstractBaseSession model and hooks providing the option
of overriding the model class used by the session store and the session
store class used by the model.

Moved the lookup in Field.swappable_setting to Apps, and added
an lru_cache to cache the results.

Refs #24743

Thanks Marten Kenbeek for the initial work on the patch. Thanks Aymeric
Augustin and Tim Graham for the review.

Thanks Sayid Munawar and Tim Graham for the report, investigation and
review.

Thanks David Nelson Adamec for the report and Tim Graham for the review.

When using a TransactionTestCase with serialized_rollback=True,
after creating the database and running its migrations (along with
emitting the post_migrate signal), the contents of the database
are serialized to _test_serialized_contents.

After the first test case, _fixture_teardown() would flush the
tables but then the post_migrate signal would be emitted and new
rows (with new PKs) would be created in the django_content_type
table. Then in any subsequent test cases in a suite,
_fixture_setup() attempts to deserialize the content of
 _test_serialized_contents, but these rows are identical to the
rows already in the database except for their PKs.  This causes an
IntegrityError due to the unique constraint in the
django_content_type table.

This change made it so that in the above scenario the post_migrate
signal is not emitted after flushing the tables, since it will be
repopulated during fixture_setup().

"Area man/woman" is confusing to people not familiar with the
conventions of American journalism (like me).

django-developers thread:
https://groups.google.com/d/msg/django-developers/11XvmVdx58w/sFrF0pL8LTgJ

Unused since efb0100e.