This serves two things -- kills annoying data offset management
which makes protobuf switch too ugly. And removes the need to
seek the pipe data image, which in turn is the prerequisity for
streaming migration.

Memory usage is still optimal, since we still use splice on dump
path and on restore stage we F_GIFT pipe buffers to the kernel.
Signed-off-by: Pavel Emelyanov <xemul@parallels.com>