NumPy -> Torch

[hang-yin]

High level changes:

• Numpy -> Torch conversion
• Enforce th.float32 whenever possible
• Simulation backend -> torch
• Rewrite transform utils with torch jit scripts
• Use math instead of torch whenever possible
• Fix robot state tests
• Remove Scipy rotation library
• Remove trimesh transformation library
• Reenable examples as tests
• Fix all examples (except attachment demo seg fault)

TODO:

• Bring performance back by getting rid of warp.synchronize() call when fetching data from annotators