Add support for approximate leave future out.

[seabbs]

At the moment the package supports using loo to produce estimates of the leave one out information criterion. This is not appropriate for most nowcasting tasks where the aim is to nowcast future data. This means that most model selection needs to be done by fitting multiple nowcasting models and evaluating the results. Doing this is computationally costly and not tractable for many users

Instead recent work on approximate LFO may be useful but this requires implementing on a model-specific basis. However, it is likely it can be implemented for at least the family of models supported in epinowcast. Potentially, some theory work is needed here as we are estimating multiple targets at once for each nowcasting vs a single time point as given in the examples below.

Resources:

https://mc-stan.org/loo/articles/loo2-lfo.html https://arxiv.org/abs/1902.06281

[seabbs]

There is a nice implementation of this here which we could draw inspiration from.

[seabbs]

There is another nice custom implementation of this in mvgam (which is a nice package more generally with lots to learn from).

[seabbs]

Pinging @zsusswein for visibility.