Split making canister calls into two steps

[timohanke]

It would be nice to first call a prepare function that returns the request id but doesn't actually send the request to the IC. Then call a commit function that sends the request.

That way we could write the request id to a journal before the request is made. The problem is that frontends may crash in the middle of making a call and then cannot find out which calls were made before the crash. With a log available they can find out at least if they come back up within 5 minutes which is often the case (not always though).