`get_laplace_approx` should provide initial guesses for `x0` in a more principled manner

[Michal-Novomestsky]

Currently, the minimize step to obtain $x^*$ which maximises $\textnormal{logp}(x \mid y, \theta)$ simply randomly initializes the initial value for the minimizer - this can be problematic if the optimum is far from this point.

Ideally, the initializer should produce a random point around some point of interest - the mean/mode of the prior on the latent $x$ is a good choice (note that $x$ is Gaussian and thus mean = mode).

This may help alleviate #568

[ricardoV94]

What about using the last optimized point?

[jessegrabowski]

What about using the last optimized point?

This is also my recommendation