Some actions should use experiment hierarchy sometimes

Created by: wilsonmr

I think that there are certain cases where we do actually want to groupby experiment. It's especially important at the moment because of the way data is loaded, but even in the future I think it would still be beneficial because it avoids doing operations like D_i * C_ij * D_j where C_ij =0 (for off block diagonal)

In particular: if fitthcovmatis None then there are no correlations outside of experiments and then constructing total covmat can be done by using block diagonals. Likewise total chi2 is just the sum of experiment chi2. Finally in filters._filter_closure_data the generation of pseudodata is faster if we generate it experiment by experiment. There might be other cases I haven't thought about.

I think with this in mind, it might be worth using explicit nodes for some actions related to these which depend on fitthcovmat which I think could solve for example #860 (closed)

Also since running some more closure tests seems imminent I foresee this potentially slowing down the process of filtering the data considerably, although I haven't check this yet.