Replace raw cudaMalloc calls with CUDACachingAllocator (#57083)
Summary:
Replace raw cudaMalloc calls with CUDACachingAllocator
Pull Request resolved: https://github.com/pytorch/pytorch/pull/57083
Reviewed By: zou3519
Differential Revision: D28058989
Pulled By: ezyang
fbshipit-source-id: 84e2d0937e3ad5e3db9ae5a5e584d8c90954e213
