xdiff: import new 32-bit version of count_masked_bytes()

xdiff: import new 32-bit version of count_masked_bytes()

Import the latest 32-bit implementation of count_masked_bytes() from
Linux (arch/x86/include/asm/word-at-a-time.h).  It's shorter and avoids
overflows and negative numbers.

This fixes test failures on 32-bit, where negative partial results had
been shifted right using the "wrong" method (logical shift right instead
of arithmetic short right).  The compiler is free to chose the method,
so it was only wrong in the sense that it didn't work as intended by us.

Reported-by: Øyvind A. Holm <sunny@sunbase.org>
Signed-off-by: Rene Scharfe <rene.scharfe@lsrfire.ath.cx>
Signed-off-by: Junio C Hamano <gitster@pobox.com>