ROLLER: Fast and Efficient Tensor Compilation for Deep Learning