Optimize the simulation smoother algorithm

[helske]

Current implementation of the simulation smoother algorithm used in walker_glm is somewhat slow, as I am not taking account the fact that we are simulating multiple samples from the same model. The corrections are relatively straightforward, and should be implemented at some point. Essentially same code is available in bssm package as well, so it should be pretty easy to just convert those codes to Stan.

[helske]

Actually, it would be better to replace the simulation smoother with psi-APF, which scales better with long time series.