How to Plant Trees in Language Models: Data and Architectural Effects on the Emergence of Syntactic Inductive Biases