x86: use LFENCE instead of CPUID before reading TSC for x86_64

According to Intel 64 and IA-32 Architectures Software
Developer’s Manual, volume 3, chapter 8.2.5, LFENCE provides
a more efficient method of controlling memory ordering than
the CPUID instruction. So use LFENCE here, as all 64-bit
CPUs have LFENCE.

Signed-off-by: Daniel Leung <daniel.leung@intel.com>