Fix online trafficlight_detection

[Berghojo]

Tasks:

• figure out why the nn with 99% accuracy during training predicts wrong in online mode

Goal:

• have a reliably working traffic light detection

[ourendingdays]

Links from Lennard und Julian

• https://github.com/ll7/paf21-1/blob/8cc49ad2cfa75a8221654d32b9db061a02d139d1/components/carla_ros_bridge/sensors.json#L25
• https://carla.readthedocs.io/en/0.9.12/ref_sensors/

[Bonifatius94]

Resons for Failure:

• colored image encoding was bgr, had to flip channels to rgb (because the nn was trained on rgb images)

Additional Changes (#62):

• image preprocessing was a bit messy, is now simplified
• neural network was simplified, now only uses ~1k weights