Improve RL training setup with multi-agent warm starts and richer collision metrics

[FatemehAB]

• Training script can now load pretrained SB3 models, use stronger multi-agent defaults, and write cleaner WandB metadata.
• Gym environment returns dict observations, keeps the simulator on the chosen device, records videos when asked, and surfaces the new collision metric.
• Simulator setup now double-checks scenario data, creates waypoint goals, and keeps background traffic controllers working together as expected.
• Single-agent wrapper shares the new observations and still returns rewards and done flags as simple numbers.