Regression in 7ec330ee.
Thanks Tim Graham for the review.

Python docs say, "it's usually not necessary to invoke these methods directly."

Thanks Tim Graham and Simon Charette for the reviews.

Per deprecation timeline.

After refs #16614, integers might appear outside the WHERE clause.

Thanks Trac alias karyon for the report and Tim for the review.

Thanks Ursidours for the report.

Thanks Maxime Lorant for the report and Anssi for the suggestion.

Thanks Loïc for the suggested improvements and Tim for the review.

Thanks to Jamie Matthews for finding and explaining the bug.

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.

Thanks Collin Anderson for the review.

There's always a better word.

This allows these tests to run on databases that reject duplicate NULL
values in unique fields e.g. SQL Server.

Forward port of 4fb50a9a from stable/1.7.x.

Thanks Anssi Kääriäinen and Tim Graham for the reviews. Refs #17001 and #22650.

Regression from f51c1f59 when using select_related then prefetch_related
on the reverse side of an O2O:

Author.objects.select_related('bio').prefetch_related('bio__books')

Thanks Aymeric Augustin for the report and tests. Refs #17001.

Thanks Claude Paroz for the report.

Regression introduced by commit 97774429. Refs #21410.

The original patch for custom prefetches didn't allow usage of custom
queryset for single valued relations (along ForeignKey or OneToOneKey).
Allowing these enables calling performance oriented queryset methods like
select_related or defer/only.

Thanks @akaariai and @timgraham for the reviews. Refs #17001.