Improved batch rpc mode

[ripienaar]

Today, and since batching was first added to mcollective, batching basically picks n nodes and complete them as a batch before moving to the next batch. This ends up looking like a jigsaw kind of pattern with some nodes fast, most about the same and some slow - sometimes very slow leading to a long tail where no/few nodes are doing anything.

With go I think we can do a bit better by using a sliding window style batching where at any time we have n machines in flight, soon as we get one reply we unlock 1 new node from being called. This way the batch is always full and the impact of slow nodes is felt in the overall call rather than each batch.

We cant change --batch to work like this so probably we will add a new style of request, something like --window or WindowsRequest() or something like that.

This will look a lot more like the old mco puppet runall in that it always keeps a set of nodes busy (though there we had to make calls inbetween to check they are done, this wont do that)