Fixed regression caused by a bug in the 32-bit strict memory access code in jdmrgss2...

Fixed regression caused by a bug in the 32-bit strict memory access code in jdmrgss2.asm (contributed by Chromium to stop valgrind from whining whenever the output buffer size was not evenly divisible by 16 bytes.)  On Linux/x86, this regression generated incorrect pixels on the right-hand side of images whose rows were not 16-byte aligned, whenever fancy upsampling was used.  This patch also enables the strict memory access code on all platforms, not just Linux (it does no harm on other platforms) and removes a couple of pcmpeqb instructions that were rendered unnecessary by r835.

git-svn-id: svn+ssh://svn.code.sf.net/p/libjpeg-turbo/code/branches/1.2.x@838 632fc199-4ca6-4c93-a231-07263d6284db