resolve-undo: teach "update-index --unresolve" to use resolve-undo info

The update-index plumbing command had a hacky --unresolve implementation
that was written back in the days when merge was the only way for users to
end up with higher stages in the index, and assumed that stage #2 must
have come from HEAD, stage #3 from MERGE_HEAD and didn't bother to compute
the stage #1 information.

There were several issues with this approach:

 - These days, merge is not the only command, and conflicts coming from
   commands like cherry-pick, "am -3", etc. cannot be recreated by looking
   at MERGE_HEAD;

 - For a conflict that came from a merge that had renames, picking up the
   same path from MERGE_HEAD and HEAD wouldn't help recreating it, either;

 - It may have been Ok not to recreate stage #1 back when it was written,
   because "diff --ours/--theirs" were the only availble ways to review
   conflicts and they don't need stage #1 information.  "diff --cc" that
   was invented much later is a lot more useful way but it needs stage #1.

We can use resolve-undo information recorded in the index extension to
solve all of these issues.

Signed-off-by: Junio C Hamano <gitster@pobox.com>

builtin-update-index.c

@@ -433,7 +433,18 @@ static int unresolve_one(const char *path)
 
 	/* See if there is such entry in the index. */
 	pos = cache_name_pos(path, namelen);
-	if (pos < 0) {
+	if (0 <= pos) {
+		/* already merged */
+		pos = unmerge_cache_entry_at(pos);
+		if (pos < active_nr) {
+			struct cache_entry *ce = active_cache[pos];
+			if (ce_stage(ce) &&
+			    ce_namelen(ce) == namelen &&
+			    !memcmp(ce->name, path, namelen))
+				return 0;
+		}
+		/* no resolve-undo information; fall back */
+	} else {
 		/* If there isn't, either it is unmerged, or
 		 * resolved as "removed" by mistake.  We do not
 		 * want to do anything in the former case.