ENH support the array_API for TRF decoder

Describe the new feature or enhancement

On a project with a student, we are relying on mne's TRF model, which is super efficient, thanks for that!

One bottleneck is that we have data embeddings on GPUs, and we would like to stay on GPU for the TRF.
With the current improvement of the array_api support on scikit-learn (see here), Ridge can now be called on GPU with the same object/API.

Would it make sense to adapt mne TRF code to allow using the object with the array API when passed torch tensors for instance?

Describe your proposed implementation

Change the way the lag view are constructed to use the array_api compat layer, and allow pass the data in the original format to the scikit-learn estimator when it is supported.

Describe possible alternatives

N/A

Additional context

No response

Start by locating the TRF decoder and the code that constructs lag views, then inspect how the array_api compatibility layer and scikit-learn estimator are used. The change is complete when supported array-API inputs such as torch tensors can remain in their original format through the TRF path and be accepted by the estimator.