Gathering improvement ideas

[rlouf]

I copy in this thread all the improvement suggestions that were made here and there.

1. In the integrators we use jax.grad to differentiate the kinetic energy. However we can compute a closed-form derivative which will be faster to evaluate, which could implement it in the gaussian_euclidean_metrics. Since the gradient of the kinetic energy is likely a lot faster to evaluate that the potential we don't expect a noticeable improvement for large models.