Softmax Output Approximation for Activation Memory-Efficient Training of Attention-based Networks