Fixed #29367 -- Fixed model state on objects with a primary key created with QuerySet.bulk_create().

COUNT is more expensive than EXISTS; use the latter when possible.

Instead of loading all QuerySet results in memory, count the number of
entries. This adds an extra query when list() or tuple() is called on the
choices (because both call __len__() then __iter__()) but uses less
memory since the QuerySet results won't be cached. In most cases, the
choices will only be iterated on, meaning that __len__() won't be called
and only one query will be executed.

Thanks Tim Graham for the fix.

Obsolete since numpy 1.12.0.

Regression in cfff2af0.

Thanks Alex Gaynor for contributing to the patch.

Fixed #29286 -- Fixed column mismatch crash with QuerySet.values() or values_list() after combining an annotated and unannotated queryset with union(), difference(), or intersection().

Regression in a0c03c62.

Thanks Tim Graham and Carlton Gibson for reviews.