Learning Patient-Specific Cancer Survival Distributions as a Sequence of Dependent Regressors