KmeansTrainer, Change the default initialization to RANDOM_NO_DUPLICATES

Today the default is RANDOM

Added enhancement label

I guess changing this would break the API compatibility? wouldn't it? ( I am not saying we shouldn't do it)

Will not break the API, but the ABI since the outputs will be different in some rare corner cases.

mentioned in issue bob.learn.libsvm#9 (closed)

mentioned in commit 0c3971fa

mentioned in merge request !24 (merged)

mentioned in commit 16c6ff5d

closed