[FEATURE] allow unrolling expert trajectories in get_model_performance

[EllingtonKirby]

Adding this as a sanity check, mainly for the nuplan data.

My assumption is that when replaying the log trajectories we would see 100% goal completion, and very low off-road and collision rates. This would indicate that "perfect" driving on these scenarios is possible (I had a slight fear that the nuplan agents had a higher off-road rate due to an issue with the data).

When I evaluate 1300 nuplan scenarios with the expert trajectories we see:

        goal_achieved_frac      collided_frac      off_road_frac      other_frac     
                      mean  std          mean  std          mean  std       mean  std
dataset                                                                              
test                 100.0  0.0           0.0  0.6           1.2  6.1        0.0  0.0
train                100.0  0.0           0.0  0.5           1.4  8.6        0.0  0.0

Agent-based metrics 

Total agents: 20918.0 in 1300 scenes
Collision rate: 0.028683431446552277
Offroad rate: 0.5975714884698391
Goal rate: 100.0
Other rate: 0.0

The off-road rate is still a bit higher than I might expect, but it is still significantly lower than the best we have recorded on nuplan scenarios so far (~8%).