Add improved support for parallelization and related graph opts (#5257)
Summary:
Pull Request resolved: https://github.com/pytorch/glow/pull/5257
- Add RescaleQuantized parallelization support to graph opts' parallelization code
- On NNPI, mirror Rescale parallelization for FC/Relus that come before it
- Sink Reshapes below Quantize and ConvertTo
- Remove unnecessary ConvertTo when following a Dequantize (i.e. just change the elem kind of the Dequantize instead)
Test Plan: Added unit tests
Reviewed By: hyuen, mjanderson09
Differential Revision: D25947824
fbshipit-source-id: 771abd36a1bc7270bf1f901d1ec6cb6d78e9fd1f
