This makes the traversal of index be in sync with the tree traversal.
When unpack_callback() is fed a set of tree entries from trees, it
inspects the name of the entry and checks if the an index entry with
the same name could be hiding behind the current index entry, and

 (1) if the name appears in the index as a leaf node, it is also
     fed to the n_way_merge() callback function;

 (2) if the name is a directory in the index, i.e. there are entries in
     that are underneath it, then nothing is fed to the n_way_merge()
     callback function;

 (3) otherwise, if the name comes before the first eligible entry in the
     index, the index entry is first unpacked alone.

When traverse_trees_recursive() descends into a subdirectory, the
cache_bottom pointer is moved to walk index entries within that directory.

All of these are omitted for diff-index, which does not even want to be
fed an index entry and a tree entry with D/F conflicts.

This fixes 3-way read-tree and exposes a bug in other parts of the system
in t6035, test #5.  The test prepares these three trees:

 O = HEAD^
    100644 blob e69de29bb2d1d6434b8b29ae775ad8c2e48c5391    a/b-2/c/d
    100644 blob e69de29bb2d1d6434b8b29ae775ad8c2e48c5391    a/b/c/d
    100644 blob e69de29bb2d1d6434b8b29ae775ad8c2e48c5391    a/x

 A = HEAD
    100644 blob e69de29bb2d1d6434b8b29ae775ad8c2e48c5391    a/b-2/c/d
    100644 blob e69de29bb2d1d6434b8b29ae775ad8c2e48c5391    a/b/c/d
    100644 blob 587be6b4c3f93f93c489c0111bba5596147a26cb    a/x

 B = master
    120000 blob a36b77384451ea1de7bd340ffca868249626bc52    a/b
    100644 blob e69de29bb2d1d6434b8b29ae775ad8c2e48c5391    a/b-2/c/d
    100644 blob e69de29bb2d1d6434b8b29ae775ad8c2e48c5391    a/x

With a clean index that matches HEAD, running

    git read-tree -m -u --aggressive $O $A $B

now yields

    120000 a36b77384451ea1de7bd340ffca868249626bc52 3       a/b
    100644 e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 0       a/b-2/c/d
    100644 e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 1       a/b/c/d
    100644 e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 2       a/b/c/d
    100644 587be6b4c3f93f93c489c0111bba5596147a26cb 0       a/x

which is correct.  "master" created "a/b" symlink that did not exist,
and removed "a/b/c/d" while HEAD did not do touch either path.

Before this series, read-tree did not notice the situation and resolved
addition of "a/b" and removal of "a/b/c/d" independently.  If A = HEAD had
another path "a/b/c/e" added, this merge should conflict but instead it
silently resolved "a/b" and then immediately overwrote it to add
"a/b/c/e", which was quite bogus.

Tests in t1012 start to work with this.
Signed-off-by: NJunio C Hamano <gitster@pobox.com>