Avoid slow compute on all ones sample weights by splitting code path.

Fix some bugs in the random state handling