Obtain num boosted rounds directly from booster.

[trivialfis]

The PR tries to remove the hack for obtaining number of boosted rounds in Python. I implemented the getter in c++ which directly counts the number of trees.

Most of the tests are already in tests/python/test_training_continuation.py, where best_ntree_limit is heavily tested.

[RAMitchell]

I'm reluctant to create an API function for a single scalar variable unless there is no alternative. Can we get this information in another way?

[trivialfis]

@RAMitchell

Can we get this information in another way?

Right now we are using len(get_dump()) and pick the parameters like num_parallel_tree and num_class in Python/R to calculate boosted rounds.

I'm reluctant to create an API function for a single scalar variable unless there is no alternative.

As #5460 being sorted out, I am thinking we give a shape attribute to the booster, like how many trees, how many output for each tree, how many groups, size of random forest etc ... Along the way we create a friendly API for specifying these attributes and remove related configuration in Learner.