I2MVFormer: Large Language Model Generated Multi-View Document Supervision for Zero-Shot Image Classification