[WIP] Use config_context for device configuration.

[trivialfis]

This is an early working-in-progress PR for using global configuration for device. There are still plenty of things I need to do, the PR is for the CI so that I can get a better picture on how many changes I need to make.

Related issue: #7308

For sklearn, we will support both the config_context and constructor parameter device (replacing the gpu_id training parameter) for easier hyper-parameter search.

xgb.XGBRegressor(device="CUDA:0", tree_method="hist") # ok, easy for parameter search with sklearn
xgb.XGBRegressor(device="CPU", tree_method="approx") # CUDA + approx will throw an error

with xgb.config_context(device="CUDA:0"):
    xgb.XGBRegressor(tree_method="hist") # ok, easier for setting global configuration.

[trivialfis]

Closing to avoid spamming the list.