FloE: On-the-Fly MoE Inference on Memory-constrained GPU