Revert to Modal backend (ZeroGPU needs PRO); keep inference.py ready

Switch inference to in-process ZeroGPU (inference.py); drop Modal calls

Deploy TinyWorld + Crisis Mode (Among-Us co-op survival)