Factor out intermediate storage

[THargreaves]

This PR completely removes intermediate storage (solving #37) and makes changes to the callback system to be compatible with this change. The main points are:

• The ancestors are now stored in the particle distribution. This doesn't change functionality at all as far as I can tell.
• The callbacks are now parameterised by a subtype of CallbackTrigger e.g. PostPredictCallback. These can be added by the user if they have a special time they want to callback. It's a bit of a pain that these need default implementations that do nothing—this is also quite annoying in that it fails silently if you get the type signature wrong in your callback definition as it just runs the default that does nothing. There's probably a cleaner way to do this but I think this is fine for this PR
• This change means we can get rid of all deep copies in the filtering code (🎉🥳) and let's us do some nice in-place operations making the predict/update cleaner. This does mean that all callbacks need to deepcopy now but they should have been doing that already. Solves Remove unnecessary deep copies #27.

The only issue this introduces is that it makes the guided filter a bit harder to implement since the old log weights are "forgotten" by the time we get to the update step. There are definitely a few options of how to handle this so not too worried for now—just a case of finding the most elegant.

Let me know if @charlesknipp @FredericWantiez have any thoughts otherwise I will merge soonish so you guys can build on this modified interface.