STen: Productive and Efficient Sparsity in PyTorch