- This allows to emit vectorized loads/stores for CUDA math intrinsics. - A few intrinsics should be lowered as CUDAMath not CUDAFastMath ones. - Fixed the code block identation.