Add uncertainty to estimates

[njtierney]

Two options (paraphrased from Nick Golding)

1. Bootstrap - bootstrap resample the polymod contact data multiple times, fit a different contact model to each, and then use each model to create a separate matrix - those will be the bootstrap samples of the matrix.

2. Conditional simulations from the fitted model (as opposed to the mean prediction with marginal noise, which would just create a bunch of speckly contact matrices and ruin the structure). Maybe we can pull some methods out of this package? It shouldn't be too difficult once we find the right objects inside the mgcv object, but will be fiddly and take time to write the code.