amd64: revamp memcmp

Borrow the trick from memset and memmove and use the scale/index/base addressing
to avoid branches.

If a mismatch is found, the routine has to calculate the difference. Make sure
there is always up to 8 bytes to inspect. This replaces the previous loop which
would operate over up to 16 bytes with an unrolled list of 8 tests.

Speed varies a lot, but this is a net win over the previous routine with probably
a lot more to gain.

Validated with glibc test suite.