how to write a fused softmax kernel in triton that eliminates redundant memory accesses and outperforms pytorch's native implementation.