[vulkan] Fix Addmm prepacking to persist after GPU flush (#48313)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/48313
Previous prepacking method would cause stored data to be lost whenever data is flushed during a `.cpu()` call. I updated the weight/bias prepacking to use the same method from `conv2d` in order to avoid this.
Test Plan: Imported from OSS
Reviewed By: AshkanAliabadi
Differential Revision: D25125405
Pulled By: SS-JIA
fbshipit-source-id: 2533994d522d90824fc25ee78c54016cfd0f3253