fix: guard model.to() calls against backend proxies in inference and validation

[imagra93]

InferenceRunner._set_device() in models/base/inference.py and GazeInferenceRunner._set_device() in models/l2cs/inference.py both call self.model.model.to(device) unconditionally. When the inner model is a _BackendEvalProxy (ONNX / TensorRT backend), the proxy has no .to() method — device management is the runtime's responsibility — so the call raises AttributeError and prediction fails.

Fix: wrap both .to() calls with a hasattr guard. model.device is still updated so the wrapper reflects the intended device; only the PyTorch-specific .to() transfer is skipped for proxies.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 621cf18c0b

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[EHxuban11]

This changes behaviour of the shared validator and thus affects every single model, which worries me. I think that such big change could come after the v1.2.0 release.

This is the pushback that codex wrote while I was reviewing the PR:

This changes behavior in the shared `BaseValidator._setup()` path, so it affects every
validator/model, not only backend proxy inference.

Specifically, validation now moves `self.model.model` to `self.device` and overwrites
`self.model.device`. That may be the right fix for native PyTorch validation device
overrides, but it is a broader behavior change than the proxy `.to()` crash this PR
describes.

I reproduced a regression with `TorchScriptBackend`: after validator setup,
`backend.device` changes from the string `"cpu"` to `torch.device("cpu")`. That then
changes later backend device comparisons/warnings, because backend code treats
`device` as a runtime string.

For v1.2.0 I’d prefer the narrow fix: guard the two inference `.to()` calls, and
either drop the shared validator mutation or split it into a follow-up PR with
backend-preservation tests.