Conv-Basis: A New Paradigm for Efficient Attention Inference and Gradient Computation in Transformers