Task-agnostic Decision Transformer for Multi-type Agent Control with Federated Split Training