Try to avoid rare bug on one ci node.

Try to avoid rare bug on one ci node.
A guess is that the win_lock call by a process may be finishing the comms posted by the raccumulate call from the other process.
And this would mean having max(0,undefined) as a result if memory is not initialized before win_lock_all.
I didn't find anything saying lock_all should be blocking/collective, so I'm not modifying it for so little.
Other guess: something's rotten with locks, as I think we don't check properly for them before posting rma calls.