LLM-Meta-SR: In-Context Learning for Evolving Selection Operators in Symbolic Regression