[SPARK-7375] [SQL] Avoid row copying in exchange when sort.serializeMapOutputs takes effect

This patch refactors the SQL `Exchange` operator's logic for determining whether map outputs need to be copied before being shuffled. As part of this change, we'll now avoid unnecessary copies in cases where sort-based shuffle operates on serialized map outputs (as in #4450 /
SPARK-4550).

This patch also includes a change to copy the input to RangePartitioner partition bounds calculation, which is necessary because this calculation buffers mutable Java objects.

<!-- Reviewable:start -->
[<img src="https://reviewable.io/review_button.png" height=40 alt="Review on Reviewable"/>](https://reviewable.io/reviews/apache/spark/5948)
<!-- Reviewable:end -->

Author: Josh Rosen <joshrosen@databricks.com>

Closes #5948 from JoshRosen/SPARK-7375 and squashes the following commits:

f305ff3 [Josh Rosen] Reduce scope of some variables in Exchange
899e1d7 [Josh Rosen] Merge remote-tracking branch 'origin/master' into SPARK-7375
6a6bfce [Josh Rosen] Fix issue related to RangePartitioning:
ad006a4 [Josh Rosen] [SPARK-7375] Avoid defensive copying in exchange operator when sort.serializeMapOutputs takes effect.