Bug - Realtime cache collision

[ADBond]

Fixes #2515.

There were occasional failings of test runs in tests/test_realtime.py, particularly in CI, that would generally be resolved on re-running.

The issue was that we used id(settings) as a cache key. In CPython this is simply the address of the object in memory, and in general is not guaranteed to be unique. When new objects are put in the same memory location as those that have been garbage-collected, we got cache collisions, resulting in the wrong SQL being returned.

We use a slightly different approach for a cache key to how we check if we have some SQL already cached, depending on the type of object:

• pathlib.Path or str (representing saved models) we just use the string itself
• dict we use a json dump (possibly after converting to serialisable types)
• for SettingsCreator we use the id still, but we also store a weak reference to the object. This won't prevent it being garbage collected, but allows us to see if the object still lives. If we find a supposèd cache hit, we check the associated weakref (which we also keep in the cache), and if it's dead (and thus we have the case of recycled id values) we delete the entry and return as though we found no hit.

The custom solutions for str and dict are mainly because these cannot be (directly) weakrefed.

You can see in this PR test runs where we see that this weakref solution is triggered, and fixes the issue (with a hack to fail the tests on a separate condition, which occurs only on that logic branch).

[RobinL]

Looks good to me, thanks!