On “Scientific Debt” in NLP: A Case for More Rigour in Language Model Pre-Training Research