CodeIF: Benchmarking the Instruction-Following Capabilities of Large Language Models for Code Generation