Securing Private Federated Learning in a Malicious Setting: A Scalable TEE-Based Approach with Client Auditing