port PDM from my adversarial codebase

[WaelDLZ]

This is my implementation of PDM from the "Parting with misconceptions" paper of D. Dauner et al.

It has a lot of parameters, and I put in the default values the one that I found to work best in PufferDrive but everyone should tune it to what it wants, there's a tradeoff in compute time vs performance (coming mostly from planning_dt and planning_horizon but also from the number of candidate trajectories you consider)

[Kait0]

The code looks fine so far.
Can you add some measurements to the PR?

Some performance metrics in an eval, for example?
And what the inference speed of PDM is (how many ms per forward pass).