add some regularization to gradientboosting warm starting

[ClimbsRocks]

we can do it ourselves! make sure that each iteration increases accuracy by more than X amount, where X is likely very small (0.0001)