Refactor blit copies and reuse scalar tensor's MPS storage in BinaryOps (#52)

* Refactor blit copies and use commitAndContinue in commitAndWait()

* Remove commitAndContinue from commitAndWait()

* [MPS] Add commitAndContinue to non blocking blits. (#51)

* Merge commitAndContinue changes from mps_master

* Don't use getMPSGraphTensorFromScalar() if tensor is already on MPS device in Binary Ops
This improves performance by preventing copy from GPU to CPU and back to GPU again

Co-authored-by: Kulin Seth <kulin_seth@apple.com>

Author: 2 people

aten/src/ATen/mps/MPSStream.h

@@ -34,8 +34,6 @@ typedef void* MTLDevice_t;
 namespace at {
 namespace mps {
 
-#define USE_MPSCOMMANDBUFFER 1
-
 //-----------------------------------------------------------------
 //  MPSStream
 //-----------------------------------------------------------------