Split up transaction chunks

Split up transaction chunks

This adds the infrastructure to prevent allocation of large byte[]s,
as when those exceed 0.5-16MiB, under G1GC they end up in humongous
object region. Not only that, but after some cut-off point, the
copying of arrays starts to dominate performance.

What we do here is ensure we always receive up to a configurable
number of bytes, defaults to 256KiB, and keep those chunks in a list.
This way we may end up with larger overhead, but that really is
neglibeble -- even a 2GiB payload would end up using only about
8K arrays.

While the input/output streams are similar to
org.apache.commons.io.output.ByteArrayOutputStream, the design here
is geared towards having the intermediate representation available
as well as devolving to a single byte[] for memory efficiency
reasons.

JIRA: CONTROLLER-1920
Change-Id: I2b79a633ebf4fdf8d68d2accc644326e30b41f22
Signed-off-by: Robert Varga <robert.varga@pantheon.tech>