The Promises and Pitfalls of Using Language Models to Measure Instruction Quality in Education