This AI Paper from CMU and Meta AI Unveils Pre-Instruction-Tuning (PIT): A Game-Changer for Training Language Models on Factual Knowledge