See #1563 (comment) for context.

This will include the following changes:

1. Expose a torch.matmul function with semantics of matmul as of PR Numpy-style broadcasting for all mathematical functions #1563.
2. Document above torch.matmul function, including broadcast behavior.
3. Document that the following functions do not broadcast:
   torch.mm
   torch.mv
   torch.bmm
   torch.dot (this function does flatten)
   torch.ger (this function does not flatten or broadcast).