Handle None in loss scaling

[tjruwase]

We have seen at least one model that returns None in a list/tuple of losses. We should gracefully handle scaling in these cases by leaving the None untouched.