STeCa: Step-level Trajectory Calibration for LLM Agent Learning