mgr/cephadm: upgrade: pull image after upgrade start, and for each host

mgr/cephadm: upgrade: pull image after upgrade start, and for each host

Make 'upgrade start' return quickly, without first pulling the image.

Pull the image once to establish the image_id.

For each host, before updating the container, ensure the local image is
up to date, and if not, pull.

If a pull returns a different image_id, restart upgrade process.  (This
could live-lock if two hosts have different container registries that
return different image ids for the same image name. :/)

Signed-off-by: Sage Weil <sage@redhat.com>