Warm up every bucket at load so the first request per size is not ~10x slower

GLiNER2.5-Decide Core ML multifunction package (L64-L512) with bucket router

initial commit