Sharper Convergence Guarantees for Asynchronous SGD for Distributed and Federated Learning