Forecasting LLM Inference Performance via Hardware-Agnostic Analytical Modeling