autoxgboost support spark dataframe/xshards

[jenniew]

autoxgboost support spark dataframe/xshards, so that it can create a pipeline with spark preprocessing without collect data to driver.

[shanyu-sys]

Autoxgboost supports distributed tuning but not distributed training. In that case, not training data but trials with different hyper-parameter sets, distributed across workers in a cluster. And each trial (a ray worker) consumes all training and validation data for fit and evaluate.

Therefore it might be a little hard to support a pipeline without collecting data to driver, since currently we use driver to broadcast all data to each worker.