Re-enable ADWIN change detection

[tms-bananaquit]

ADWIN is "famous" in the literature as a change detector for arbitrary real-valued numbers. The recent signature change sets it up s.t. it can only monitor accuracy.

The X, y_true, y_pred signature will be indeterminate in any ensemble drift detector; it's not as simple as checking whether X is specified and switching modes. Seemingly the smallest change at this point is to add an argument to the init (smells bad) or to make parallel classes that handle both cases. The docstrings will be redundant, as we have in kdq-tree/HDM.