Q.deconstruct() is nondeterministic if the Q has multiple kwargs

[Harro]

Here is a branch with a randomly failing test that proves this: ​https://github.com/hvdklauw/django/blob/bug/q_destruct/tests/queries/test_q.py#L63

The biggest issue is that now makemigrations is detecting changes since we upgraded to django 2.0 (only with python 3.5) that aren't changes at all, just reordered kwargs on the Q objects in out limit_choices_to on some foreignkeys.

This also means we randomly can't commit/release because we have pre-commit hooks and CI that runs makemigrations --check --dryrun

Upgrading to python 3.6 would fix it (because of ordered kwargs) but as Ubuntu 16.04 is still the latest LTS, python 3.5 is what we are stuck with.

This is the commit that broke it: ​https://github.com/django/django/commit/508b5debfb16843a8443ebac82c1fb91f15da687

[Simon Charette]

Apart from preserving the kwargs nature of some children I'd argue that Q.deconstruct should also use django.utils.models.Q as path instead of django.db.models.query_utils.Q like we do with django.db.models.fields.* and should avoid adding _connector if it's 'AND' and _negated if it's False to kwargs because these are the default values.

[Tim Graham]

​PR

I don't understand why {'_connector': 'AND'} should be omitted but I implemented it.

[Tim Graham <timograham@…>]

In b95c49c9:

Fixed #29125 -- Made Q.deconstruct() deterministic with multiple keyword arguments.

[Tim Graham <timograham@…>]

In 9ba3df8:

Refs #29125 -- Made Q.deconstruct() omit 'query_utils' in the path and _connector='AND' since it's a default value.

[Tim Graham <timograham@…>]

In aeb35548:

[2.0.x] Fixed #29125 -- Made Q.deconstruct() deterministic with multiple keyword arguments.

Backport of b95c49c954e3b75678bb258e9fb2ec30d0d960bb from master

[Tim Graham <timograham@…>]

In fd18345e:

[2.0.x] Refs #29125 -- Made Q.deconstruct() omit 'query_utils' in the path and _connector='AND' since it's a default value.

Backport of 9ba3df82402e7e23b353da20aea6894935241ef9 from master