[Fix] Removed the possibility to create an Obs from data on several replica

[s-kuberski]

This pull request addresses #256:

It removes the possibility to create an Obs from data that is based on several ensembles. Whereas the feature is needed internally, it can be considered dangerous for users as the behavior of the function might be unexpected. To allow for internal creation of such Obs in intermediate steps, the creation is still allowed if means are passed to the function.

Further changes that I introduced along the way because they are coupled to the same problem. It is possible that I have overlooked something. In this case, these changes can be discussed and reverted.

• It is not possible anymore to reweight an Obs that already contains several ensembles. The only setting where I see that this could be a valid procedure is the case where algorithmic parameters that do not affect the reweighting but only the autocorrelation time have been changed such that two independent streams are not real replica and instead treated as separate ensembles. Also in this case, I would propose to force the user to reweight before combining the observables.
• It is not possible anymore to correlate two observables that are based on several ensembles. I think that this is a feature that has to be applied to primary observables instead of derived observables, because otherwise averages are not properly defined.
• The function merge_obs has been adapted to the same rules. The function header is extended to better explain what is done here.

Apart form this, the biggest changes in the code are in the JSON and dobs I/O, where I had to adapt to the new way to create the Obs and in the tests, where we had used the illegal creation in some cases.

It would be an option to add a routine that combines observables from several pseudo-replica by averaging them, either weighted by the number of configs or by the number of independent configs.

[fjosw]

Looks good to me!

[fjosw]

Maybe @JanNeuendorf can confirm that this actually addresses the issue he raised.

[JanNeuendorf]

I think that it solves the issue because I see no remaining way to get a multi-ensemble Obs without taking an explicit average.

[s-kuberski]

By extending the tests to less trivial cases in some cases, I've discovered a small bug that I had introduced before. This is now fixed and I believe that all changes are now covered by the tests.