Mutiprocessing could be faster

[dongxijia]

Thanks for your codes, it really helps.
And I notices the 60 line code in AdaBoost.py:
votes[i, :] = np.array(list(pool.map(partial(_get_feature_vote, image=images[i]), features)))
It maybe accurs error because pool.map would mix the order of feature,
while the 89 line in AdaBoost.py usd the feature as below, but features' order may differ from votes[i, ;], :
best_feature = features[best_feature_idx]

I noticed using pool.map(partial(_cal_feature, features), image_faces) and pool.map(partial(_cal_feature, features), image_no_faces) can handle this problem and make full use of cpu. The training time will reduce from 30s to 8s in my 6 CPU compute.
The details you can see in my Pull requests.