amd64: switch csum_partial_copy_generic() to new calling conventions

... and fold handling of misaligned case into it.

Implementation note: we stash the "will we need to rol8 the sum in the end"
flag into the MSB of %rcx (the lower 32 bits are used for length); the rest
is pretty straightforward.

Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>