refactor read_tarball to share tree construction logic

[StefanKarpinski]

Multiple usages of read_tarball now effectively want a tree representation of the tarball structure—tree hashing & symlink copying, in particular. Currently symlink copying doesn't build a tree and instead does a lot of splitting and joining on slashes, which is annoying and inefficient. The tree hashing code builds a tree for itself, but this same approach could probably be shared, simplifying both use cases.