Vectorized/GPU Sim Evaluation / ManiSkill 3 Port

[StoneT2000]

• Add GPU sim environments for bridge dataset (in ms3 repo: [Feature] Real2Sim Eval Digital Twins haosulab/ManiSkill#536)
• Modify Octo to perform batched inference
• Modify Octo to accept torch tensors/dl packed jax arrays
• Modify RT1-X to do batched inference

Closes #35 #36

[StoneT2000]

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!