Overhaul state/particle/group subset from history

[richfitz]

How this is 700 lines I don't know. It was fiddly though.

This PR looks at extracting state from the filters with dust_likelihood_last_history and dust_likelihood_last_state. There are two related problems.

• Previously these accepted index_group to subset by group. But this was problematic because it depended on the index_group passed into dust_likelihood_run()! I have removed the index_group argument from these and they return the groups that we ran with originally which is pretty much always what is wanted, and more advanced use will be rare enough that we can either implement it when asked or people can subset manually.
• I've moved where index_state goes from filter creation to filter running. This requires a small dance so that we can ensure that the filter is created before we validate the index but that in turn tidies up my least favourite part of the monty initialisation. We might export the little "ensure initialised" function I have written there later. This keeps the two index arguments together which is nice.

The PR is large because:

• this affects a bunch of generated code
• we use the history saving mechanism for simulate
• there's some similar-but-different code for the filter/unfilter and the discrete/continuous splits

[weshinsley]

OK - I think I sort of understand this one, but would probably be good to talk through. The monty bits will probably need tweaking after that other PR - not sure what you'd like to do first so I won't merge