Derivatives with respect to parameters

[dpole]

1. Agnostic about the structure of kwargs (which is the same of the
   output derivatives of the models): It can be a nested structure.
2. To address it, _apply applies a function to the leaf of nested
   dictionaries and lists
3. Drafted derivatives for CorrelatedFactorizedCrossSpectrum, Join
4. Derivatives wrt to multipoles are compressed from (ell, ..., ell)
   to (ell, ..., 1). This saves a lot of memory and operations, but
   may lead to mistakes. We have to decide whether it is worthwhile
5. Relatively minor changes here and there which, however, might break
   some compatibility with external codes
6. Only the variables should be passed to diff (i.e., those that are
   None in the defaults) and returned in the output dictionary/list

[beringueb]

Minor fixes in frequency.py and power.py. Still need to look at cross.py.
I'm not sure how the changes you've implemented in 4) will propagate when used in FactorizedCrossSpectrum.
At first glance, it seems the implementation in FactorizedCrossSpectrum would depend on the cls used, which I'm not sure is something we want. But I need to have a further look. Will do asap !

[beringueb]

Had to revert to previous implementation of 4) for now. I have tested everything else and it's working as before. Need to complete the implementation of CorrelateFactorizesCrossSpectrum.
One thing I noticed that might be worth thinking about:
Even when the SED is fixed, a diff method still needs to be implemented for (if we want to use it in FactorizedCrossSpectrum). There is probably a simple way around that (eg. check whether any defaults parameters is None?)