The `remove()` and `clear()` methods of the related managers created by
`ForeignKey`, `GenericForeignKey`, and `ManyToManyField` suffered from a
number of issues. Some operations ran multiple data modifying queries without
wrapping them in a transaction, and some operations didn't respect default
filtering when it was present (i.e. when the default manager on the related
model implemented a custom `get_queryset()`).

Fixing the issues introduced some backward incompatible changes:

- The implementation of `remove()` for `ForeignKey` related managers changed
  from a series of `Model.save()` calls to a single `QuerySet.update()` call.
  The change means that `pre_save` and `post_save` signals aren't called anymore.

- The `remove()` and `clear()` methods for `GenericForeignKey` related
  managers now perform bulk delete so `Model.delete()` isn't called anymore.

- The `remove()` and `clear()` methods for `ManyToManyField` related
  managers perform nested queries when filtering is involved, which may
  or may not be an issue depending on the database and the data itself.

Refs. #3871, #21174.

Thanks Anssi Kääriäinen and Tim Graham for the reviews.

Application template now includes an empty migrations module, and
the autodetector will only make initial migrations for apps with
empty modules.

Fixed #21499 -- Migrations won't work if field signature changes

Thanks Gunnar Scherf for the report and the suggested patch.

Fixed #21512 -- Added more complete information about Pillow and PIL in model_fields and model_forms tests.

Updated ModelAdmin to use form.instance when passing parent model to
child inlines for add_view. There is effectively no change in the
change_view since the previously passed 'obj' is the same as form.instance.

Thanks to meshy for report, and EvilDMP and timo for review.

Since Python 2.5, KeyboardInterrupt and SystemExit are not subclasses of
Exception, so explicitly reraising them before the “except Exception” clause
is not necessary anymore.

A test for annotations incorrectly assumed that the first instance
(in the test) of a model using AutoField for PK will always get pk=1.
The test was changed to compare against actual instance id instead.

Failure was introduced with test in [19e5cd77]

Thanks glarrain for the report.

The settings reference documentation doesn't seem the right place to
invite users to write their own db and cache backends.

Also the actual wording makes the task sound trivial, which is hardly
the case; writing a custom db backend is a very difficult task, and
writing a custom cache backend is full of gotcha.

This only affects the new cache api and not the deprecated get_cache.

Refs #21012

Thanks to trac user paultag for the report.

This (hopefully) ensures that the cache are created the same way as before
the offending commit.