Make log_lik methods correct by obtaining latent parameter draws

[athowes]

In PR #259 we added a slightly incorrect version of the log_lik_latent_lognormal function. This allowed progress on model comparison in the Ebola vignette (#48).

It would be good to make this correct however.

For background see:

• Function for pointwise log likelihood #240
• A thread on the Stan forum summarising the issue

In short I see the two most promissing routes as:

1. Try to put the required pwindow_raw and swindow_raw draws into some part of the prep object
2. Use ... to pass in the fit object then extract the draws

Future investigation required to figure out how to do this!

[seabbs]

An easy way to do this is to refactor the approach to latent variables so that pwindow and window are defined via non-linear formulas as in the original vignette - that will in turn mean they are available as part of the prep object

[athowes]

Fixed effect by case_id requires having case in the data (which it isn't right now).