Consider defaulting to fast logrank in a future release

[erikcs]

This option (#1509) would obviate the need for rounding/discretization to speed up survival forest training as this algo has constant-time scaling in the number of time points (can remove notes in the docstring + warning/help messages on discretizing times).

Storing the final predictions can take up alot of memory if the time grid is very large, so still useful to be mindful of.

If we can make a breaking change, then maybe it's worth considering a lower default alpha for survival_forest with low events, something like min(0.05, mean(D) / 2.5))