Harden GPU fallbacks for previews and conditioning

Remove API-layer hardcoded CUDA envmap device

Fallback image conditioning to CPU on unsupported GPUs

Make attention backend selection configurable

Fallback RMBG preprocessing to CPU on unsupported GPUs

Reduce cold start with snapshot caching and warmup

Fallback to basic GLB export on CuMesh failure

Keep TRELLIS resident on GPU by default

Support gated HF DINOv3 downloads

Add compiler toolchain for Triton runtime builds

Delay TRELLIS GPU imports until first request

Initial Docker API Space for TRELLIS.2