The current interface used to fetch shuffle data is not very efficient for
large buffers; it requires the receiver to buffer the entirety of the
contents being downloaded in memory before processing the data.

To use the network library to transfer large files (such as those that
can be added using SparkContext addJar / addFile), this change adds a
more efficient way of downloding data, by streaming the data and feeding
it to a callback as data arrives.

This is achieved by a custom frame decoder that replaces the current netty
one; this decoder allows entering a mode where framing is skipped and data
is instead provided directly to a callback. The existing netty classes
(ByteToMessageDecoder and LengthFieldBasedFrameDecoder) could not be reused
since their semantics do not allow for the interception approach the new
decoder uses.

Author: Marcelo Vanzin <vanzin@cloudera.com>

Closes #9206 from vanzin/SPARK-11235.