add censor set and KM time sampler.

[juAlberge]

different things added:

• add a censor set to separate the training of the CIFS from the censoring distribution
• add an argument: time_sampler.
  • time_sampler='uniform' -> stay as it was. We sample a time uniformly between (0, t_max)
  • time_sampler='kaplan_meier' -> we train a Kaplan Meier estimator to learn the survival distribution and sample time to have more sampled time where the events of interest actually happen.

TODO before PR:

• Adding documentation
• Adding tests

[Vincent-Maladiere]

I 100% agree to have this at some stage, but after a discussion with @GaelVaroquaux, we thought that we needed to wait slightly before creating this wrapper, as it could make the wrong message to the community. Does this idea still hold, @GaelVaroquaux?