Decentralized Learning for Overparameterized Problems: A Multi-Agent Kernel Approximation Approach