zlib: Fix big performance regression

zlib: Fix big performance regression

Commit 340fdf1303dc ("zlib: Port fix for CVE-2016-9841 to U-Boot")
brings a big performance regression in inflate_fast(), which leads
to watchdog timer reset on powerpc 8xx.

It looks like that commit does more than what it describe, it
especially removed an important optimisation that was doing copies
using halfwords instead of bytes. That unexpected change multiplied
by almost 4 the time spent in inflate_fast() and increased by 40%
the overall time needed to uncompress linux kernel image.

So partially revert that commit but keep post incrementation as it
is the initial purpose of said commit.

[trini: Combine assorted patches in to this one, just restoring the
        performance commit]

Fixes: 340fdf1303dc ("zlib: Port fix for CVE-2016-9841 to U-Boot")
Signed-off-by: Christophe Leroy <christophe.leroy@csgroup.eu>
Acked-by: Michal Simek <michal.simek@amd.com>
Signed-off-by: Tom Rini <trini@konsulko.com>