findandmodify should eventually support scatter/gather operations

as it does a single document (at most), mongos presumably would serialize messages to each shard. it would try one, and if a result, return it. if no result, it would try the next shard. for best performance, on a new findandmodify command request, it would make sense for the mongos to start with the shard that last returned a result, and only advance once it has nothing to do.