Dynamic Privacy Allocation for Locally Differentially Private Federated Learning with Composite Objectives