This is a performance tweak. Since inverse diagonal operators are automatically
converted to a pure diagonal operator with the inverse on its diagonal, this
commit saves memory.

Furthermore, the zero_mode keyword is gone for AmplitudeOperator. It is not used
in the demos and is not really appropriate from a theoretical point of view
anyways. We will add in a better option later.

commenting getting_started_3.py a bit more, tuning parameters to be more stable (if exp nonlinearity is used)

results.png: make KL.position and MOCK_POSITION power spectra lines
thicker so they are clearly identifiable.

loop.png: do not overwrite the intermediary loop plot, save one
for each iteration

Sample more often to make better use of the computation time.
Beforehand, when starting from zero, convergence was far from
reached in the time the demo ran, while much time was used
to optimize for the initial samples.