Towards Foundation Models for Scientific Machine Learning: Characterizing Scaling and Transfer Behavior