Refactor to drop Stan

[seabbs]

As we are just optimising do we want to refactor to drop Stan as it can be a bother? Slow would ve r fast but equally a bother would ve c++

[sbfnk]

This makes me wonder why we're not sampling weights and then propagate the uncertainty in the ensemble.

[seabbs]

To get a weight posterior? I think we could potentially do that but that does go beyond the original stacking inspiration