CUDA: Accelerate MXFP4 table lookup using `__byte_perm` (#15451) #11

Job	Run time
macOS-arm64	2m 42s
ubuntu-22-vulkan	7m 32s
macOS-x64	6m 2s
ubuntu-22-cpu (x64, ubuntu-22.04)	3m 48s
windows (vulkan, x64, -DGGML_VULKAN=ON, ggml-vulkan)	4m 14s
windows (opencl-adreno, arm64, -G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-window...	1m 30s
windows-cpu (arm64)	2m 29s
windows-cpu (x64)	5m 27s
ios-xcode-build	11m 46s
windows-cuda (12.4)	45m 1s
windows-sycl	10m 56s
windows-hip (radeon, gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032)	45m 2s
release	1m 36s
	2h 28m 5s

Provide feedback