Regression hidden by 6fe7eb60.

This reverts commit 6fe7eb60.
These checks are still in use outside of memcpy() optimizations.
Note this revert breaks the build on some platforms due to code rot.

There is no need to test OS support for SSE2, and later. If plain SSE
works, the OS preserves the XMM registers and that's all we care.

 - Assembler code out of line
 - ARM NEON run-time detection
 - Better choice of registers
 - Prefetching
 - ARMv6 SIMD optimizations where Advanced SIMD not available

Scheduling is not completely optimal.

In retrospect, I should have done that first. Oh well...

macosx: fixed NSRangeException when triggering modifier keys without characters as enabled by some international keyboards (close #7268)

If the sample per packet is wrong, denote to qt_version 0

Ref: http://forum.videolan.org/viewtopic.php?f=14&t=103013

This accounts for sporadic assert failed at vlc termination.

This will be used later.

Hardware floating point must be enabled explicitly.

This also removes the NEON_CFLAGS substition variable and the check for
ARMv6, neither of which are anymore not used.

minizip should not and cannot be compiled without zlib.

extras/package: iOS: enable thumb for compilation, clean up the AS definitions and disable freetype for now

Should be tested on Android/Linux.

Thanks to Mans Rullgard for the help.