APOLLO: A Simple Approach for Adaptive Pretraining of Language Models for Logical Reasoning