Enhance multi-modal support

[ehsk]

Goal: Reproduce previously produced reward curves on ChartQA

Old	New (this PR)
achieved more than 80 after around 30 steps	same results and fixed crash after 40 steps with before (orange) vs after (green)

• streams=redis (changing to files substantially slows down the actor as the size of rollouts explode), added max_stream_size to control the size of Redis' queue.
• freeze_vision_encoder option added. (not used for the above results though)
• Minor improvements in launch.py
• Batch size can be greater than 1:

x-axis=optimizer steps	x-axis=time