Analyzing Methods for Efficient Language Model Adaptation with Domain-Specific Selective Layer Expansion