In histogram GBDTs, the parent node negative loss should be computed only once

[NicolasHug]

In _find_best_bin_to_split_helper(), it'd be more efficient to only compute the (negative) loss of the parent once, instead of at each bin.

In other terms, we should compute
negative_loss(sum_gradients, sum_hessians, l2_regularization)

before the for loop.

link to relevant code

I don't expect this to be a terrific improvement though.

Marking this easy, but I would not recommend as a first issue since it involves cython.

[fhoang7]

Taking a look at this...was having trouble finding a test suite around the cython file though.