Add back z-scoreing for flow matching

[manuelgloeckler]

🚀 Feature Request

This was diabled because naive z-scoreing as input layer can have very negative consequence performancewise. This however can be done similar to how it is done in NPSE with access to the time-dependent moments.

For the simple FM we have implemented with $\mu_0=0$ and $\sigma_0=1$, $\mu_1 = \text{theta.mean(0))}$ and $\sigma_1 = \text{theta.std(0)}$ this should be given by
$$\mu_t = t\mu_1$$
and
$$\sigma_t = \sqrt{t^2 \sigma_1^2 + (1-t)^2 }$$

Z-scoreing can then be doen automatically as for all the other estimators.