Add Gradient Boosting Incidence

[Vincent-Maladiere]

What does this PR implement

This is the implementation of GradientBoostingIncidence. This also introduces a bug fix for y_binary, detailed in #10.

TODO:

• [x] the documentation is partial and needs to be completed (Olivier thinks the docstring of the public class is good enough for now). Some narrative guide will follow in a follow-up PR once we have several estimators to compare.
• [x] the tests are missing
• [x] write an example (e.g. check that marginal estimation is unbiased with various level of censoring),
• [x] add 0-time negative targets explicitly to bias the model towards starting the CIF curves at (0, 0).
• [x] make the monotonic constraint optional.

Remarks I replaced GradientBoostedCIF from the tutorial notebook with this PR and obtained the same results, both for the estimator and the brier score metric.

[ogrisel]

I think this is good enough for a first iteration (assuming CI is green).