Improve initialisation for GMM likelihood

[DeaglanBartlett]

Currently the optimisation (not the MCMC) appears unreliable if starting far from the true parameters for the GMM case (when the data is drawn from a GMM of known parameters). This can probably be improved by better initialisation. Perhaps fit a GMM to the xobs values directly without doing the function fit and use these parameters as an initial guess for the full optimisation problem.

[DeaglanBartlett]

Now initialise full GMM optimisation with a GMM fit to just the xobs values, using sklearn's GaussianMixture.