The original git_indexer_commit resolved REF_DELTA objects by retrying the
entire unresolved list on each pass — O(N^2) for packs with deep or wide delta
chains.

There are 2 dependency maps:
 - `desp_by_offset` - base object pack offset to `OFS_DELTA` children
 - `deps_by_oid` - OID to its `REF_DELTA` children

Delta resolution then becomes a single BFS traversal: each resolved object
immediately looks up and resolves its children in O(1), giving O(N) total
work. Thin pack fixup (injecting missing bases from the ODB) was preserved and
integrated into the same traversal.

New regression test (`test_pack_indexer__delta_chain`) was added to cover
out-of-order delta resolution. Existing tests pass.

I also verified the change by cloning a giant repo (~5.7 Gb) before and after
with some clone example of a bare mirror of a single branch and compared the
resulting directories using `diff -rq`. They are identical.

The change doesn't make too much difference with default `-DUSE_SHA1`,
but when I switch to use `CommonCrypto` on MacOS. I get:
```
8:57.53 total
```
instead of previous
```
17:12.05 total
```