get.max_acc fails to calculate correct accuracy on edge case

[mmauri]

Hi Laurae, Congrats with this superb library. With few observations xgboost some times returns the same probability to each one, and having the same probability with different labels makes get.max_acc report the inverse of accuracy & the max threshold also is incorrect if we assume that positive obs are > threshold (not >=)

To reproduce it: get.max_acc(rep(0.502,8),c(0,1,1,1,1,1,1,1))

It gives 0.125 accuracy with 0.502 threshold

Thanks for your time