Reduce texgen resolution to 1024 and enable CPU offload to lower GPU memory pressure

Add explicit runtime diagnostics and exception logging for GPU execution

Prefetch Hunyuan weights at startup to keep GPU task within duration budget

Load torch before custom_rasterizer import to resolve libc10 dependency

Try torch 2.5.1 + tencent custom rasterizer wheel; simplify rasterizer bootstrap

Build custom_rasterizer from source at runtime with discovered CUDA paths

Fix runtime CUDA shared library path and lazy imports

Initial Space: image+GLB attire transfer with Hunyuan3D-2 texgen (ZeroGPU 60s)