Similar to #912, but for the verbose argument of Parallel.

With sklearn, some models use Parallel without supporting verbose (e.g., OneVsRestClassifier). Exposing the verbosity level in the context manager would help deal with such cases.