Adding support for input data on GPU in BayesSearchCV

[ghzut]

Hi,
first of all thank you very much for reviving this package. I'm mainly using BayesSearchCV to optimize a XGBoost BDT but this might affect other parts of the package too. Since XGBoost 2.0 it is possible to inplace_predict, if the data is stored on the same device the booster uses to train. Because of the immense speed up, the device in my case is a gpu. It would be nice, if cupy arrays would be accepted as an input of the fit() function or alternatively an option that allows to convert to cupy, so the hyperparameter optimization can also benefit from this.

Cheers
ghzut