Towards Automated Functional Equation Proving: A Benchmark Dataset and A Domain-Specific In-Context Agent