Smarter sampling of messages to reply to

[andreaskoepf]

Message trees are "grown" randomly. This is fine for the first two levels if the maximum branching factor is something like <= 5 but for deeper trees it generates to many "lonely" messages with only one reply-variant. For ranking we need multiple replies per message.

Proposed solution: When a tree reached a certain percentage of its final size (e.g. 50%) the sampling is no longer purely random but replies/children are primarily inserted where other siblings exist.