Compute likelihood via particle filter

[richfitz]

Merge after #8, contains those commits

We'll get better statistical tests on this later, but some poking about with it suggests we're correct :-/

For the future tests, we have a model where we derived the likelihood a different way and show it converges. But for that I need to implement support for compiling models as we go, and that's a few PRs away.

This PR implements the same particle filter algorithm as used in dust1 and mcstate1:

1. advance the model up to the next data point
2. compare particles to the data
3. convert the log likelihoods into a set of weights (this requires moving out of the log basis into a normal basis, which is half of the gross bookkeeping - scale_log_weights)
4. resample using the "systematic sampling" scheme (this is the other half of gross bookkeeping, see resample_weight - tbh I forget how this works but the R reference implementation in test-filter-details.R is actually very simple). This samples, with replacement, particles in proportion to the weight computed in step 3
5. accumulate the log likelihood which is the log of the average likelihood (not the average log likelihood) over each step

Some future features are flagged in the code

• early exit for filter invocations that will not be accepted
• save trajectories as the filter proceeds
• save snapshots at some time points in the simulation

The RNG seeding is complex, and different to the way that we do this in dust1:

• every filter gets its own rng, with the aim that a filter that computes multiple parameter sets at once could split apart. This frees us up to parallelise in different ways (e.g., openmp, mpi or something else). This means for n particles and m parameter sets we need (n + 1) * m streams
• we will pull these from the rng state by jumping so that the first n + 1 streams go to the first parameter sets, and of that the first stream goes to the filter and the remaining n go to the model
• I've written a state extraction routine which checks this, and a test that shows you get the same answer running a filter with multiple parameter sets as a set of filters seeded appropriately

[weshinsley]

Talked through, and think I get it...