Text this: Rethinking I/O Caching for Large Language Model Inference on Resource-Constrained Mobile Platforms.