[cuDNN][cuDNN V8 API] Improve hot path heuristics performance in V8 (#90811)
Small optimization for the hot path when thrashing the cache with dynamic shapes; in most cases we don't need the fallback generator so we can omit it unless needed later.
CC @ptrblck @ngimel
Pull Request resolved: https://github.com/pytorch/pytorch/pull/90811
Approved by: https://github.com/ngimel
