Parallel Runner has worse performance at same timestep/episode than Episode Runner

[PMatthaei]

Reason:

• Steppers did not read correct metrics f.e. reward was read correctly if policy team id set to 0 but in the build plan the policy team was at position 1. So done array was wrong accessed reading the done boolean of the scripted agent
• learning seems to diverge or be very fluctuating in parallel steppers
• episode stepper achieves better results in same time
• check if logging is wrong and policy actually is good