Add MODE switch: LOCAL (on-device, default) vs MODAL (GPU endpoints)

Modal-backed Space: run models on Modal GPU endpoints, Space is a thin HTTP client

ZeroGPU: move models to CUDA inside @gpu calls, not in the main process

ZeroGPU: read models from module globals, not pickled args

Integrate ZeroGPU Space path with card-cache, sampling, and diversity fixes

finished base