Created by: FrostML

Add register op_data_type of pad2d. In this PR, float/double/int/int64_t are all support for pad2d op. According to #21942 , atomicAdd() for double-precision floating-point numbers is not available on devices with compute capability lower than 6.0. This PR fix this problem by using platform::CudaAtomicAdd() rather than atomicAdd().