Add consistent checks for arguments of coregistration classes

[rhugonnet]

Not a big priority, but right now inputs of coregistration classes are inconsistently checked, which can yield in some cases non-user friendly error messages if they are wrongly defined (even though the type hinting should provide some clarity for the user when that is the case).

We should add checks similar to the _check_inputs_bin_before_fit used only in NuthKaab in affine (for now), and the checks in BiasCorr.__init__ which check the "fit_or_bin" parameters, for all types of input.
Currently the subsample argument is checked but only once it runs through fit(), while it would be better to have an error raised straight at class instantiation (=when it is defined). And random_state is not checked at all.