I suspect the reason the nullid is in the index in the last position
is that it lets index[i] for regular revision number, even when index
was just a regular Python list. An alternative solution would have
been to reserve revision number 0 for the null revision. I don't know
why that wasn't done. Now that we have classes backing the index, we
can easily make index[-1] get the nullid without having to put it last
in the list and including it in the len().
This patch just hides the nullid -- it will still be accessible at
index[len(index)].
I realize that this will be annoying when checking out across this
commit for debugging (including bisection).