CoLD: Counterfactually-Guided Length Debiasing for Process Reward Models