Simplify conv2d weight prepacking (>2x pipeline-creation speedup) (#3368)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/3368

SS-JIA has previously written two implementations of convolution weights prepacking for CPU (before and after [PyTorch PR #84973](https://github.com/pytorch/pytorch/pull/84973)). Originally, I translated the second implementation to GPU since it is more readable. Now, I translate the first implementation to GPU and switch to it since it requires less steps.

Reviewed By: SS-JIA

Differential Revision: D56617129

fbshipit-source-id: b84533a45e1daf31f7fdec6707a518f3de57ce4c