Selective Induction Heads: How Transformers Select Causal Structures In Context