Use ZeroGPU for Space inference

[vvennela]

Motivation

• Ensure model loading and inference run inside Hugging Face ZeroGPU allocations on Spaces so GPU work is limited to allocated calls and local/test runs keep working without the Spaces SDK.
• Provide a small compatibility layer so spaces.GPU can be applied where available while remaining a no-op in local/test environments.

Description

• Add zerogpu.py, a compatibility helper that applies spaces.GPU when present and provides an identity decorator otherwise.
• Wrap Diffusers image generation and pipeline loading with the ZeroGPU decorator by updating art.py to use @gpu for DiffusersImageClient.create_art and DiffusersImageClient.load.
• Wrap Transformers-backed LLM load/complete paths by adding @gpu to Nemotron and MiniCPM load/complete methods in local_llm.py.
• Add spaces to requirements.txt and update README.md to suggest zero-a10g hardware and document the ZeroGPU usage.

Testing

• Running pytest -q in the default environment failed during collection due to module import path resolution errors (automated run reported import errors).
• Running the test suite with the repo on PYTHONPATH via PYTHONPATH=. pytest -q succeeded and reported 258 passed.

---

Codex Task

[bolt-new-by-stackblitz]

Run & review this pull request in StackBlitz Codeflow.