s390/checksum: make ip_fast_csum() faster

Convert ip_fast_csum() so it doesn't call csum_partial(), but instead
open code the checksum calculation. The problem with csum_partial() is
that it makes use of the cksm instruction, which has high startup
costs and therefore is only very fast if used on larger memory
regions.

IPv4 headers however are small in size (5-16 32-bit words). The open
coded variant calculates the checksum in ~30% of the time compared to
the old variant (z14, march=z196).

Signed-off-by: Heiko Carstens <hca@linux.ibm.com>
Signed-off-by: Vasily Gorbik <gor@linux.ibm.com>