Long-Short History of Gradients is All You Need: Detecting Malicious and Unreliable Clients in Federated Learning