Potential bug in FullSubnetLaplace.sample()

[elcorto]

Looking at #268, I think the same issue is present in FullSubnetLaplace.sample() since it uses the same construct samples @ self.posterior_scale, which was inspired by FullLaplace.sample in #216. It might also affect other sample methods, such as the one in DiagSubnetLaplace, but I'd need to inspect that code more closely.

In any case I think before fixing this, one needs tests catching the issue. My first guess is to draw (potentially many) samples, calculate the empirical cov matrix and compare against posterior_covariance, as @2bys and @simoneiriksson have suggested.