PT_BREAK: Implement multiple dispatch

[ailzhang]

pytorch/pytorch#25653

[dlibenzi]

We get called into copy_() where self is not an XLA tensor?!?

Input tensor is not an XLA tensor: CPUFloatType

What are the new copy_() semantics?