Move calculations away from GPU in Bandwidth profilers (#4445)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/4445

This is simply to have a more accurate result when doing bandwidth profiling by removing calculations that can be done outside the shader, leaving only the read operations behind.

Reviewed By: copyrightly

Differential Revision: D60396870

fbshipit-source-id: 53dc774f08b1284b7e1eff0cd14ac06c01c7532a