mpi/amd64: use SSE2 for shifting instead of MMX

* mpi/amd64/mpih-lshift.S: Convert to SSE2.
* mpi/amd64/mpih-rshift.S: Ditto.

On current Intel processors, MMX instructions is slower than SSE2. Switch
lshift and rshift functions to use SSE2 registers instead of MMX.

• Signed-off-by: Jussi Kivilinna <jussi.kivilinna@iki.fi>