Improve pre-shuffle aggregation performance in mrTriplets

[rxin]

I realized today that we can actually save the position of the source vertex and the target vertex in the array in EdgeTriplet, and then in aggregation, we can simply use that position to update the aggregation value (without actually doing any hash lookups).

Issue 101 will make this easier to implement.

https://github.com/amplab/graphx/issues/101