Efficient Sparse Training with Structured Dropout