Bias Runs Deep: Implicit Reasoning Biases in Persona-Assigned LLMs