Generalize prediction cache.

[trivialfis]

Generalize prediction cache for other uses.

• Extract most of the functionality into DMatrixCache.
• Limit the size of the prediction cache to 32 items. This should be plenty. When the number is exceeded, it's usually because the user forgets to free the input DMatrix during inference, and by limiting the size of the cache, we are actually improving performance for those cases.
• Move API entry struct to an independent file to reduce dependency on the predictor.h file.
• Add a test.

I'm working on learning to rank related changes. One hurdle I run into is how to cache sorted indexes inside Metric. Currently, the cox metric adds a member function to meta info, while the AUC has an ad-hoc cache. This PR proposes we reuse the prediction cache in these places.

[trivialfis]

@hcho3 Thank you for the review, all comments are addressed.