#304 LambdaMART uses metric cutoff which could be 0 for MAP

[Alex Libov]

When training a lambdaMART model, and using the MAP metric, the default "k" for MAP is 0.
LambdaMART uses the k as its cutoff which severely hinders performance when the cutoff is 0.
0 as a cutoff is not treated by lambdaMART as "all".