allow arbitrary cross validation fold indices

[owlas]

• use training indices passed to folds parameter in training.cv
• update doc string

xgboost cv allows user to pass scikitlearn split objects or pass a list of indices for each fold. Indices are of the form:

[ (fold_1_train_idx, fold_1_test_idx), (fold_2_train_idx, fold_2_test_idx) ]

The current implementation uses the test indices to form dtest and forms dtrain from all indices that are not in the test indices.

An improvement would be to use the passed train indices explicitly (i.e. form dtrain from the train indices). This allows completely arbitrary cross validation strategies.