add more flexibility with delays

[sbfnk]

This adds support for

• fixed delays (mean only), or fixed delay distributions (mean + sd)
• fixed truncation distributions (mean + sd)
  as well as unit tests for these.

It is based on #304 but separated out for clarity.

[seabbs]

See #308 for an alternative approach where delays are first estimated in R vs in model. This should be more efficient in theory but is incomplete.

[sbfnk]

Precomputing the PMFs is of course a good idea to improve performance. This now contains the changes from #311 and #312 but implements the precomputation in the transformed data block of the stan model.

Benefits compared to the approach in #312 are that it avoids bleeding/duplication of model code into R and supports a mixture of fixed and variable delays. Downside is that it adds stan code and thus makes the stan model harder to read.

Passes checks locally but needs some more thorough checking that it hasn't introduced a bug in the model itself. Also the indexing of generation time and delay PMF needs review - I think as implemented it bases the discretised generation times at 1 and delays/truncation at 0.

[seabbs]

This in principle looks good though a little confused where everything is coming from. Did this draw code from the convolve_pmf branch or roll it's own implementation. If it's own I need to get my head around if it's handling convolving pmts correctly.

Code from convolve_pmf for reference.

https://github.com/epiforecasts/EpiNow2/blob/convolve_pmf/inst/stan/functions/convolve.stan

[sbfnk]

It includes the improved discrete convolution from #311 - not sure how that relates to the convolve_pmf. It doesn't contain anything from convolve_pmf explicitly.

[Bisaloo]

Sorry, I was just trying to get more familiar with EpiNow2 and the current development and ended up making some comments 🙈