FastTree Gradient of Logistic Loss prohibits small learning rates

[rogancarr]

Issue

The gradient of the Logistic Loss implemented in FastTree uses the LambdaRank-style sigmoid parameter, set to the learning rate. This quashes the gradients for small learning rates. While this works well for classification tasks, when used by the General Additive Model (GAM) trainer, it prohibits learning with small learning rates. However, the GAM learning-by-boosting technique implemented here requires small learning rates to be stable.