Bring over Penny's code to use INLA space-time model as the level 1 stacker

[goldingn]

Work in progress:

• [x] switch countries to those that are contiguous and with sufficient data
• [x] port over mesh creation
• [x] replace linear regression stacker with (blank) INLA functions
• [x] refactor functions to do fitting and prediction in one go
• [x] plumb the L0 model predictions into the L1 model
• [x] convert the data and L0 OOS and in sample predicitons into INLA stacks in fit_predict_level_one_model.R
• [x] extract and return predictions from the fitted INLA model
• [x] update code around final predictions
• [x] debug inevitable problems

[goldingn]

This now runs all the way through to making the final predictions, and makes multi-band prediction for multiple insecticides.

There are still various things to check and improve, and the predictions have weird vertical stripes which suggest some kind of issue with either the spatial model or the covariates.

[goldingn]

vertical stripes were due to longitude (and latitude, and other things) being included as level 0 covariates, now removed

[goldingn]

This runs through for me, and I've confirmed (by making L0 models run on only the intercept) that the spatial GP bits are running and doing something remotely sensible