Deep Active Inference Agents for Delayed and Long-Horizon Environments