This reverts refactoring done in 13224ea4
that introduces a performance regression for NFS when reading files that
don't exist. open() forces a cache invalidation on NFS, while stat()ing a
file just uses the cache and is very quick.

To give a specific example, say you have a repo with a thousand packed
refs. Before this change, looking up every single one ould incur a thousand
slow open() calls. With this change, it's a thousand fast stat() calls.