perf(glb): further reduce bake params to avoid 120s timeout

perf(glb): speed up bake by reducing hardcoded defaults

revert: disable NeAR main-process preload; restore torchvision weights=...

fix: enable NeAR CPU preload by bypassing torchvision weights CUDA init

fix(zerogpu): lazy-import nvdiffrast in postprocessing_utils

fix(near): defer renderer/tone-mapper init to first GPU ensure

Space app: SLaT session cache, tone-mapper safety, CUDA/inference fixes

revert: remove sparse attention _linear inference-tensor clone hack

fix(sparse-attn): clone inference feats before Linear when grad is enabled

fix(sparse): spconv replace() must wire new feats into data.features

fix(export): clone slat feats instead of inference_mode on full export

fix(export): inference_mode for export_glb_from_slat (inference tensor slat)

fix(dino): drop default HF near-assets; use torch.hub facebookresearch/dinov2

fix(spaces): Hdri torch.load after ConvNeXt; default off CPU preload with spaces

fix(load): CPU-only checkpoint load for ZeroGPU main process

fix(pipeline): do not treat ckpts/... as Hugging Face org/model id

fix(space): split ZeroGPU cold-start paths and add timing diagnostics

fix(pipeline): respect pretrained rembg (RMBG-2.0); optional NEAR_REMBG_MODEL

fix(pipeline): force BiRefNet to ZhengPeng7; drop gated rembg fallback

fix(pipeline): fallback rembg when gated RMBG-2.0 lacks HF auth

HF Space: drop ./vox2seq from requirements; vendored vox2seq_pytorch fallback

restore: full Space tree + assets (recover from minimal force-push); keep ZeroGPU app.py