Hard link detection is now done in a separate pass on the in-memory tree,
and duplicates can be 'removed' and 're-added' on the fly. When making any
changes in the tree, all hard links are re-added before the operation and
removed again afterwards.

While this guarantees that all hard link information is correct, it does
have a few drawbacks. I can currently think of two:

 1. It's not the most efficient way to do it, and may be quite slow on
    large trees. Will have to do some benchmarks later to see whether
    it is anything to be concerned about.

 2. The first encountered item is considered as 'counted' and all items
    encountered after that are considered as 'duplicate'. Because the
    order in which we traverse the tree doesn't always have to be the
    same, the items that will be considered as 'duplicate' can vary with
    each deletion or re-calculation. This might cause confusion for
    people who aren't aware of how hard links work.