[Kait0]

Large PR that fixes the pedestrian collisions in CARLA by introducing pedestrians into training.
Also teleportation in nuPlan and CARLA is replaced by letting the policy directly control the respective vehicle model of the two simulators.

The final configuration was tested on 4 models with 2 different mini-batch sizes, each trained with 1 8-GPU node for 4 days. Results are consistently strong and reasonably stable across training seeds.
Eval:

https://forge.coreweave.com/wandb/emerge_/nightly-multi-long/runs/k_scaled_0073_2000?nw=nwuserbern_jaeger
https://forge.coreweave.com/wandb/emerge_/nightly-multi-long/runs/k_scaled_0074_2000?nw=nwuserbern_jaeger
https://forge.coreweave.com/wandb/emerge_/nightly-multi-long/runs/k_scaled_0074_1000?nw=nwuserbern_jaeger
https://forge.coreweave.com/wandb/emerge_/nightly-multi-long/runs/k_scaled_0073_1000?nw=nwuserbern_jaeger

Other changes:

• model now waits in front of the red light instead of ontop of it. To do this moving the entire vehicle across a red light is considered an infraction now + some margin behind and in front. To let the model continue to succeed at traffic light we now input the remaining yellow light to the policy and increase the minimum yellow time to 2.0.
  In eval harnesses where this information is not available we estimate it by using the minimum yellow time - observed yellow time.

• Some lanes in nuPlan were significantly outside of the training distribution. We clip lane observations now at eval time to the max observed in training.

• Changed self-play eval per episode steps to match GIGAFLOW

• Changed the map to have the red light position match the CARLA traffic light trigger

• In nuPlan and CARLA clip the policy to have no reversing only stopping and forward driving.

• Added GIGAFLOW style policy unrolling to waypoints + using the nuPlan controller. Results were not good, off by default.

• nuPlan Static clutter filter now tests box overlap with lane polygons via proximal map objects instead of a centre-point test.

• fix for partner stopped time in nuPlan

• The planner no longer pickles its policy and env into every simulation log, reduced nuPlan eval log size from (28 MB → 0.2 MB) per step and total eval from (45 GB → ~0.1 GB).

• no custom slot encoder when eval on CPU

• OpenCV threads pinned to 1

• longest6 partner slots to 40

• pedestrians expempt from stopped-time zeroing

• Fix broken wandb eval results logging and replace it with a dedicated script eval_wandb_report.py

• Shard self-play eval more fine-grained per GPU instead of per node. Increase number of eval scenarios

• float16 rollout buffer to safe memory