Resume fetch after failure

[alban]

Initially reported on rkt/rkt#1280:

If a fetch is interrupted by a crash or by power loss, it currently must completely start over the download. It would be great if it at least stored layers, if not resumeable bytes like a basic program like wget.

[runcom]

this may be possible with the docker v2 registry only (since it supports the Range header when pulling layers). I can make the change needed to support this (when registry is v2) if ppl are interested.