- Add the label to the saved prior
- Remove obsolete write function

Fixes one issue in #75

Moritz Huebner: made minimum/maximum np.nan by default and simplified the assignment self.minimum/self.maximum

Removed some obsolete methods

Previously, `parameters` in `fill_prior` was only ever used to fill in
the non-standard parameters if sampling in something other than the
defaults. This directy checks that. The advantage is a cleaner logic and
we no longer assume the likelihood has a
`non_standard_sampling_parameter_keys` attribute.

Adds a lnprob method to a few of the priors. Need to add this for all
priors and a method to calculate it, if not given.

The samplers and prior access the parameters via
`likelihood.waveform_generator.parameters`. But, this is the only way
that the waveform_generator was accessed outside of the likelihood. We
simplify this by insisting that the Likelihood has the attribute
parameters (which is set, in this case, from the waveform_generator).

- Add attribute parameters to likelihood
- Remove all look-ups of waveform_generator outside of the likelihood.
  This means the waveform_generator is a likelihood-specific thing!

- Removes the 'bad_keys' part of the logic. I believe this was
  essentially adding the keys which where not given in the prior and
  that had no default prior, and then removing them at the end.
- Change the logging output slightly: now the only warning that is given
  is when no default prior is known. Other output at info level
- Add in what the waveform_generator knows about missing keys (useful
  for diagnosis
- Formatting of the docstring

- Rename n_samples to size to fit with standard numpy conventions
- Fix check to handle multiple values

Functionally, this changes nothing, but it makes it clear that the
filled priors are the dictionary `priors`.

This reverts commit f04ab987