Kernel Self-Attention in Deep Multiple Instance Learning