[cuDNN][cuDNN V8 API] Use suggest memory format for cuDNN V8 API (#87617)
Fixes some failures we observed in `functorch` tests which seemed to stem from benchmark cache collisions on the same memory format. Changing the memory format to be dependent on both input and weight seems to resolve them.
CC @crcrpar @ptrblck
cc @csarofeen @ptrblck @xwang233
Pull Request resolved: https://github.com/pytorch/pytorch/pull/87617
Approved by: https://github.com/ngimel
