No easy way to test (would have to actually trigger some routing
events from kernel) but the code is clearly wrong as written.
If the header says there is a submessage, we need to at least
skip over its bytes, not just continue to the next iteration.

Fixes #8203.

LGTM=r
R=r
CC=golang-codereviews, mikioh.mikioh, p
https://golang.org/cl/164140044