Patch autoencoder_kl_triposg: fix embedder fp32 upcast in _decode

Fix dtype mismatch in flash_extract_geometry: force fp16 for VAE decode

Fix TripoSG NaN cascade and remove stale autocast wrapper

Skip DiffDMC (CUDA arch mismatch on A10G), use marching cubes with adaptive iso-level and bg_fill

Fix NaN handling in flash_extract_geometry: preserve NaN for DiffDMC, use -1.0 for MC fallback

Pre-convert grid_logits to CPU before DiffDMC attempt (prevents CUDA ctx corruption in fallback)

Add explicit error logging + last-resort MC fallback in flash_extract_geometry

Patch flash_extract_geometry: marching cubes fallback when DiffDMC unavailable (fixes NoneType on mesh output)

Patch TripoSG flash_extract_geometry: guard empty next_points (IndexError on min())

Patch PSHuman attn: use kwargs in Attention.__init__ (diffusers added kv_heads param, breaking positional args)

Patch PSHuman unet: _load_state_dict_into_model moved to model_loading_utils in diffusers>=0.28

Fix PSHuman CLIPFeatureExtractor removed in transformers>=5.0

Replace runtime TripoSG string-patching with pre-patched files in patches/triposg/