If commit timestamps are off, we're more likely to hit a traversal
where the first path ends up traversing past the root commit of the tree.
If that happens, it's possible that the loop will complete before the second
path marks some of those final parents. This fix keeps track of the root
nodes that are encountered in the traversal, and verify that they are
properly marked.

In the best case, with accurate timestamps, the traversal will continue
to terminate when all the commits are STALE (parents of a merge-base), as
it did before. In the worst case, where one path makes a complete traversal
past a root commit, we will continue the loop until the root commit itself
is marked.

Ahead-behind count is still a valid operation, even if the two
commits don't have a common merge-base. The old implementation was
buggy, so it returned ENOTFOUND. Fixed now.

Integrate mark_parents() with the ahead_behind() code.

To be used as a basis for a function which marks nodes with parents
up to the merge base.

Indexer speedup with large objects

tree: relax the filemode parser

There are many different broken filemodes in the wild so we need to
protect against them and give something useful up the chain. Don't
fail when reading a tree from the ODB but normalize the mode as best
we can.

As 664 is no longer a mode that we consider to be valid and gets
normalized to 644, we can stop accepting it in the treebuilder. The
library won't expose it to the user, so any invalid modes are a bug.