Issues in predicting on ellpack.

[trivialfis]

We implemented predicting on ellpack for reducing memory usage. But after some more testing, I found the implementation is not thoroughly considered. Here are a few more cases we need to put more thoughts on:

• What if the user change tree method but use the same DMatrix?
• What if the user train for another round but at the second round they put the training DMatrix as a validation DMatrix?

@RAMitchell

[trivialfis]

The proper solution should be waiting for the new DMatrix interface, so that we only use ellpack prediction when there's no SparsePage.