Make all Slerp operations use PyTorch rather than Numpy for simplicity

[humzaiqbal]

Hi! I used the SLERP algorithm but noticed that it first converts all torch tensors to Numpy and then does the algorithm and converts everything back. Since people would presumably be loading checkpoints which are torch by default I figure it would be best to make the code default to torch. It also cleans up some of the code, like getting rid of the maybe_torch function. Let me know if this looks good to merge, if not happy to iterate!