Annotate across merge

[Matt Mackall]

I've adapted the annotate code to work across merges. This is good,
because it means other people's names actually show up.

The new algorithm caches the results of previous annotations that it's
going to need again and thus doesn't have to recalculate everything
back to the root when it follows the second branch of a merge. So it
requires only O(edges) diff operations and a fairly small cache of
previous annotations (fork points that have been calculated and are
referenced by uncalculated nodes).

-- 
Mathematics is the supreme nostalgia of our time.