some notes on changes

• removed tf image resize in favor of jax image resize which permits vmapping and does not have a issue where tf will allocate a lot of GPU memory if given a jax array instead of numpy.
• 72 parallel envs + model inference uses about 18-20 GB of GPU memory. For the envs with max 60 step episodes this completes 72 episodes of octo-small evaluation in about 1 minute vs 8 minutes (depending on GPU). 8x speed improvement!