Speed up the training (#9278)

* Move the "-expand" and "-cast" ops backward

* Hard-coding for stable diffusion,
maximize overlaps

* Use op_tyep_name instead of visual string

* Change transfer nodes to tributary nodes

* Rename tributary to overlap

* Prepare to test different decide parameters

* Prepare to print and test

* {7, 5} seems to be one of the best as before

* Find the best straighten mode 973 for stable diffusion

* Put cpu nodes into overlap node list

* Disable overlap between cpu and gpu if no cpu nodes

* Update API

* Remove magical number

* Update comment

* Remove std log message

* Remove debug code

* Static analysis

* Variable op still have activation time in cpu

* Rename (address comment)