If you are sent a paper to referee that is not publicly available, I believe it is your duty to provide your report and then pretend you have never seen the paper, at least until it appears somewhere public. You shouldn't be thinking about proving any conjectures made in the paper or generalizing its results. You shouldn't be contemplating collaborations to extend the work. The author has chosen not to make his or her results public yet, and has the right to have this privacy maintained throughout the editorial process. So I think contacting the author in situation (2) of the original post is highly inappropriate.