[cuDNN][cuDNN V8 API] Fix incorrect use of `emplace` in the benchmark cache (#97838)
`emplace` does not overwrite the existing mapped value in a map if it already exists, which can lead to repeated execution of a plan that e.g., tries to allocate an OOM-inducing workspace size and retriggers either a heuristic run (or worse, a benchmark run).
CC @ptrblck @ngimel @Fuzzkatt @syed-ahmed
Pull Request resolved: https://github.com/pytorch/pytorch/pull/97838
Approved by: https://github.com/ngimel