[Feature Request] - Tests redesign

Feature Description

For now it takes extremely long time to compile test wrappers. Making this procedure faster can significantly boost our developing progress.

This is likely caused by compiling every possibly used templates for all tasks at the same time, while only small parts of them are actually used. Meanwhile, the use of torch lib is not necessary. We should separate wrappers and only compile ones that are used, and we may also want to write our own ground truth functions for each kernel to test.