metrics.py

The new implementation, even if it looks very naive, reduces the
computation time of fpr/tpr vectors.
roc_curve computation time depends now only on the length of y_score.
For comparison, here are the results between the old and the new
implementation for the following vectors:

- 10^6 length vector  (y_score has 1000 unique values):
    - old impl.: 28.29 seconds
    - new impl.: 3.14 seconds

- 10^6 length vector (y_score has 10000 unique values):
    - old impl.: 267.61 seconds
    - new impl.: 3.64 seconds