ghash-x86*.pl: fix performance numbers for Core2, as it turned out
previous ones were "tainted" by variable clock frequency.

ghash-x86_64.pl: minimize stack frame usage.
ghash-x86.pl: modulo-scheduling MMX loop in respect to input vector
results in up to 10% performance improvement.

change.

subject to change...

Submitted by: "Noszticzius, Istvan" <inoszticzius@rightnow.com>

Don't clear the output buffer: ciphers should correctly the same input
and output buffers.

rely on indirect call to block functions, they are not as fast as
non-consolidated routines. However, performance loss(*) is within
measurement error and consolidation advantages are considered to
outweigh it.

(*) actually one can observe performance *improvement* on e.g.
    CBC benchmarks thanks to optimization, which also becomes
    shared among ciphers.