Consider moving from broadcasting to weekly to daily increments in underlying latent processes

[SamuelBrand1]

We've been guided by the common approach of weekly (i.e. 7 time step) broadcasting for $R_t$, and done the same for DirectInfections and ExpGrowth infection generating processes. However, they make a bit less sense in those contexts IMO.

For consistency we should consider moving all latent models used in the pipeline to not having broadcasting.

The obvious downside is that this mean approximately x7 parameters to do inference upon, and maybe a switch to reverse-mode AD?

[SamuelBrand1]

In f2f discussion, we are considering doing this for some models in a first pass.

[SamuelBrand1]

It looks like #415 could be resolving #340 i.e. numerical instability with compiled tape AD. Conditional on this, that increases the upside of this change by diminishing the down side of swapping to reverse mode (e.g. we can use compiled tape to be faster).