Refactor DFSSerializer to remove code duplication.

[mihaimaruseac]

Summary

There is a duplication in the directory traversal between DFSSerializer and FilesSerializer. Since the later supports parallel hashing, let's prepare to use only that.

We make FilesSerializer be an abstract parent class that performs the directory traversal. We introduce ManifestSerializer for the old FilesSerializer class that was creating a manifest out of the model. We rename DFSSerializer to DigestSerializer for consistency.

Since FilesSerializer (the directory traversal) only considers files now, we need to add a _FileDigestTree class to transform the list of files and their hashes (the FileManifestItem) to a directory traversal tree, so we can build the digest for DigestSerializer in a bottom-up fashion, like before. We could have just included only the files, instead of the directory, but that would require changing a lot of expected constants in the tests. So, we add this transformation now, we plan to migrate tests to goldens and then maybe change the hashing to only include the files when rolling up to a single digest.

We still had to update one test: since the hashes are computed only for files, we no longer differentiate between a model with an empty directory and a model where that empty directory is completely removed. This is a corner case and it is ok to do this.

In fact, ignoring empty directories is part of the optimization hinted at in #197.

Release Note

NONE

Documentation

NONE