Fixed a bug with tracking our active contexts

[ayshih]

#7437 added a warning when both the assume_spherical_screen() (now, SphericalScreen()) and propagate_with_solar_surface() context managers were used together, but there is a bug in the implementation of tracking active contexts that gets confused if a given context manager is used multiple times in a nested structure, e.g.:

with SphericalScreen(...), propagate_with_solar_surface():
    with propagate_with_solar_surface():
        pass

    # Context tracker thinks that the first propagate_with_solar_surface() is no longer active
    # because the second propagate_with_solar_surface() has exited
    ...

This PR overhauls the internal tracking of context managers to use the data structure of a stack instead of an unordered set (which was peculiarly implemented as a dictionary of True/False values). Now the tracking respects the nesting of contexts and can handle repeated contexts. Plus, the contexts are stored as fully qualified names, which should avoid potential collisions.

[ayshih]

While conceptually this can/should be backported to 6.0/6.1, it does change quite a bit with private behavior, so I'm fine with keeping it to 7.0. The likelihood of this bug affecting users in the wild is low, and the consequence is simply that a warning is not being emitted when it should be.

[nabobalis]

Codecov says no test? I am not sure this is worth testing but I leave that up to you.

[ayshih]

A test feels a little unpalatable to write because this condition should never be satisfied with normal code. Someone would need to intentionally modify the tracking registry directly while within contexts, and if someone is going to do that, basically all bets are off.