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

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

finished base

Build Tabras card duel prototype