Built-in method to do "one sample from the posterior".

[casperwberg]

Effective sampling without the need to invert the hessian. Automatic standardization and decorrelation would be nice (Gaussian processes only). Nice if it would work for both process error and observation error residuals ( the latter might demand that the extra modifications of the template, e.g. require residuals to be REPORT'ed). Option to use MCMC instead of Gaussian approximation would be nice also.