Improving task difficulty modelling for robot teaming in multi-dimensional contexts: with applications in performance prediction and LLM-driven multi-robot planning