CUDA: fix MMQ writeback for int8 tensor cores (#8100) #14077

Job	Run time
Push Docker image to Docker Hub (light, .devops/llama-cli.Dockerfile, linux/amd64,linux/arm64)	10m 27s
Push Docker image to Docker Hub (server, .devops/llama-server.Dockerfile, linux/amd64,linux/arm64)	13m 15s
Push Docker image to Docker Hub (full, .devops/full.Dockerfile, linux/amd64,linux/arm64)	49m 43s
Push Docker image to Docker Hub (light-cuda, .devops/llama-cli-cuda.Dockerfile, linux/amd64)	1h 9m 56s
Push Docker image to Docker Hub (server-cuda, .devops/llama-server-cuda.Dockerfile, linux/amd64)	36m 4s
Push Docker image to Docker Hub (full-cuda, .devops/full-cuda.Dockerfile, linux/amd64)	1h 38m 30s
Push Docker image to Docker Hub (light-rocm, .devops/llama-cli-rocm.Dockerfile, linux/amd64,linux...	36m 50s
Push Docker image to Docker Hub (server-rocm, .devops/llama-server-rocm.Dockerfile, linux/amd64,l...	39m 27s
Push Docker image to Docker Hub (light-intel, .devops/llama-cli-intel.Dockerfile, linux/amd64)	12m 25s
Push Docker image to Docker Hub (server-intel, .devops/llama-server-intel.Dockerfile, linux/amd64)	15m 0s
	6h 21m 37s

Provide feedback