Never go further than the previously fetched HEADs when updating a repository

This avoids a very expensive, and quite slow, recursive search, when updating a
big repository such as linux.git. The downside of this is that we download more
data, but that's not really a problem.