Restore Gradio ZeroGPU backend with 250k input cap

Let HF Gradio runner serve mounted OpenAI API app

Expose OpenAI-compatible API with 250k context and tool-call shim

Add key mapping for SuperGemma checkpoint prefix

Force Gemma4ForCausalLM for SuperGemma checkpoint key layout

Use tokenizer fallback with Gemma4 conditional model

Use Gemma4 conditional AutoProcessor/AutoModel class

Fix Gemma4 tokenizer output handling

Add traceback diagnostics for ZeroGPU generation spike

Create experimental SuperGemma4 ZeroGPU spike

initial commit