The format stress tester prepares 10% of all transactions.

Transaction prepare locks out transaction begin in format (see WT-3849), which is going to lead to less thread interactions, and specifically less snapshot-isolation repeatable read testing in the presence of interacting threads.

Format is our primary tester of repeatable reads and is only a smoke test of the prepare functionality, I'm more concerned with the former, and less concerned with the latter.

Add a prepare configuration to format, at a low probability, so most runs don't configure prepare at all.