PERF predict_proba is slow when n_jobs > 1 for random forests

[tomMoral]

The use of n_jobs > 1 for small batch can slow down the prediction for forest models. This is probably due to the overhead incurred by using joblib (create thread, check system info, ...) which is dominate the runtime compared to the computations. This was reported originally in joblib/joblib#982.

A couple of ideas to solve this:

• Set n_jobs=1 when the size of the batch is small.
• Introduce a n_jobs_predict parameters that would default to 1/n_jobs for forests but that can be set separately.