Eyes on VLM: Benchmarking Gaze Following and Social Gaze Prediction in Vision Language Models