A unique_together is a list of tuples of fields, or a list of lists of

fields. When saving these to a signature, we normalize them to a list of

tuples, to ease comparison, but there were still comparison issues for

data not yet stored.
We now normalize whenever the property is set (either in the constructor

or when applying a unique_together value), ensuring we're comparing

the same types of data, and ensuring we're working with copies of the

original data. That also allows us to remove a bunch of deepcopy()

calls in other places.
Some changes were also made around the _unique_together_applied flag.

First off, while still a private variable, there's now a public function

(apply_unique_together()) that will set it, so that the mutation class

doesn't have to reach into private variables.
More importantly, comparison now works differently.

ModelSignature.__eq__ calls has_unique_together_changed, which only

considered the old model signature's value for

_unique_together_applied. The problem is, if objects are compared in

the other direction (which can and does happen), the wrong result will

be returned.
That's been updated to instead compare the old/new flags directly, which

is more correct and avoids an issue where evolutions get skipped if the

only change to a model is a unique_together baseline change.

Unit tests pass on all versions of Python and Django.