Spark Transformer: Reactivating Sparsity in FFN and Attention