Update and expand on the descrition of the mass matrix adaptation

[aseyboldt]

The current description is outdated and pretty minimal anyway:

This crate mostly follows the implementation of NUTS in Stan and PyMC, only tuning of mass matrix and step size differs: In a first window we sample using the identity as mass matrix and adapt the step size using the normal dual averaging algorithm. After discard_window draws we start computing a diagonal mass matrix using an exponentially decaying estimate for sqrt(sample_var / grad_var). After 2 * discard_window draws we switch to the entimated mass mass_matrix and keep adapting it live until stop_tune_at.