GPU compatible (L-)C2ST

[JuliaLinhart]

Is your feature request related to a problem? Please describe.

When working with high-dimensional datasets, computing C2ST (and particularly L-C2ST) metrics can be quite slow. This is primarily due to the reliance on scikit-learn (sklearn) classifiers, which do not support GPU acceleration. As a result, performance is limited, especially for large datasets or complex models.

Describe the solution you'd like

Introducing support for classifiers from GPU-compatible libraries, such as skorch, would significantly improve performance. By leveraging GPU acceleration, training and evaluation of classifiers for C2ST metrics could become substantially faster, enhancing both development and experimentation workflows.

Describe alternatives you've considered

Currently, no viable alternatives are available since the limitation stems from the lack of GPU support in sklearn. Expanding support to include GPU-compatible libraries appears to be the most effective solution.

Additional context

• Would require adding another library as a dependency. Need to check if this might conflict with our current dependencies.
• Adapt current implementation here
• Think about similar accelerate similar metrics i.e. here