[cuda] Fix the incorrect types in int8_gemm (#107895)
Fixes #107671
From cublas team: alpha and beta need to be of the same C++ type as of scaleType, which is `int32_t` here.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107895
Approved by: https://github.com/Skylion007, https://github.com/cpuhrsch
