[ZeroGPU] CUDA placement note

[cbensimon]

Agents (also some human devs) reading https://huggingface.co/docs/hub/spaces-zerogpu tend to lazy-load models inside @spaces.GPU (despite the example code snippet), probably out of caution. Let's make it more explicit

---

Note

Low Risk
Low risk documentation-only change that clarifies where models should be moved to CUDA and why lazy-loading inside @spaces.GPU is inefficient.

Overview
Clarifies ZeroGPU guidance around @spaces.GPU by adding a prominent note that the decorator is effect-free outside ZeroGPU environments.

Adds a new Model loading section explicitly instructing users to move models to cuda at module import time (not inside @spaces.GPU) and explains that this works via CUDA emulation outside the decorated function, with real CUDA used inside it.

^{Reviewed by Cursor Bugbot for commit d98ed6b. Bugbot is set up for automated code reviews on this repo. Configure here.}

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[pcuenca]

Big +1 to the clarifications, I've seen this happen in the wild a few times.