First Release of DiffTRe algorithm

This is the first release of the DiffTRe algorithm. It represents the code state used in the paper "Learning neural network potentials from experimental data via Differentiable Trajectory Reweighting".