Globally Optimize Job-based TypedPipe.write calls

[johnynek]

Before this, in 0.18-develop we were calling the optimizer after each write. This not only makes a sub-optimal optimization, but also for long jobs with many writes at different stages makes optimization an O(N^2) operation (times how much it costs, which is significant, to optimize each set of TypedPipes).

This pattern comes up in summingbird, so it is a practical issue for us when trying to use scalding 0.18 with summingbird at Stripe.

[ianoc-stripe]

looks like a pretty straightforward win

lgtm