Feature request - please support weight per label per sample

[tlkhm]

Currently DMatrix weight argument is per sample, the shape is [n_samples].

Please allow different weights to be specified for each sample and each label, the weight shape would be [n_samples,n_labels].

This would be useful in some multi-output binary classification problems where label importance are different for each sample.

[milannnnn]

This feature would also be useful for multi-output regression as well.