Select mvn_method when running fitted_samples()

[stefgehrig]

While post_draws() contains the possibility to set mvn_mehod (the package/method to simulate form the multivariate normal implied by the fitted vcov matrix), this option is not available in fitted_samples.gam(), which calls post_draws().

Is there a reson for this? It would be handy to switch to mvn_method = "mgcv" from the default mvn_method = "mvnfast" when using method = "gaussian" with fitted_samples.gam(). In my case, the cholesky decomposition used by mvnfast failed, so I needed to switch to the (slower) implementation.

A very easy fix would be to add the default mvn_method = "mvnfast" explicitly to the signature of fitted_samples.gam(), and the signature mvn_method = mvn_method to the call to post_draws() within fitted_samples.gam.

[gavinsimpson]

Quite right: weirdly, one of the other methods for models fitted by scam() did allow this and I think I must have gotten distracted after adding it to that method and never got round to adding to all the other methods as well.

I'm just checking some fixes for this, which should be on GitHub soon.