Text2midi-InferAlign: Improving Symbolic Music Generation with Inference-Time Alignment