Split load/train: CPU model load + GPU training separately

Fix: increase GPU timeout to 180s, remove torch_dtype kwarg

Restructure: single GPU call for load+train, no nesting

Fix ZeroGPU: load model on CPU, move to GPU inside decorated fns

Head-level membrane v2 — decompose attention by individual heads

Upgrade to gpt2-large (774M, 36 layers, 20 heads)

Upgrade to gpt2-medium (355M params, 24 layers)

Remove show_copy_button — not in Gradio 6

Restore full PyTorch demo with ZeroGPU support

Remove PyTorch dependency — demo uses PoC layers directly

Use Gradio 6.10.0, explicit server bind

Fix HF Space: CPU-only torch, lazy imports

Condensate PoC — all 4 layers + HF Spaces demo