Handling previous states with GPU-RBPF

[THargreaves]

Depending on which textbook you follow, there are various definitions of Rao-Blackwellised dynamics. Some have the inner transition model conditioned on x_{t-1}, some x{t}. Some have inner observation model conditioned on x_{t-1}, some x{t}.

My initial thought was to include both of these cases by passing prev_outer, new_outer into the inner filter. The trouble is, since we update the outer state first, this requires extra storage to remember the previous outer state.

We still need to decide how much we want to care about allocations but if we do decide we want to be sparing, we would need to change how we do this.

@charlesknipp had a great suggestion that we only pass in the new_outer state and in the rare cases the user wants to use x_{t-1} they can create an augmented state containing both (x_{t-1}, x_t). In fact, the only case I can see someone wanting to do this is for our unit test comparing to the Kalman filter.