Add AcceleratedSurvivalSplittingRule

[erikcs]

Add faster splitting rule class AcceleratedSurvivalSplittingRule that computes an approximation of the logrank criterion in SurvivalSplittingRule, but which can be updated in constant time for every split point (a short note with the computational details is here). This works essentially exactly the same as SurvivalSplittingRule, except that the numerator of the criterion is updated efficiently, while the denominator of the criterion is replaced with a simple approximation that is also updated efficiently.

The statistical performance is practically the same, but prediction output may differ slightly compared to exact criterion. For backwards compatibility, a new argument fast.logrank defaulting to FALSE is added to survival_forest and causal_survival_forest.

[erikcs]

Pinging @sebp in case this is of interest: This simple change updates the log-rank criterion used for splitting in random survival forests. The speed gains are large for data with many events, with negligible loss of accuracy. A short note is here.