Created by: Sand3r-

Once the quantization warmup is perfromed using fp32 operators, some of the cached MKL-DNN primitvies are incorrectly reused as int8 operators. This PR clears device context cache after the execution of warmup, so that the aformentioned doesn't occur.