The file header format is a bit more complex than I first realized,
especially around when a path to `data/` is included, and which size
field determines its file name length.
This type is similar to an `Either` between a `Murmur64` hash and a
`String`. This is necessary to be able to retain hash information where
the hash is not in the dictionary, but at the same time allow string
names where they are available.
Up until now, when reading a bundle, all hashes would be converted to
strings, which made sense for displaying those names. But when writing
the same bundle back, those strings ended up being re-hashed, resulting
in incorrect hashes.
Clippy's suggestion failed me, as `slice::take` only yields _at most_
the given number of elements, but `Vec::with_capacity` doesn't resize
in a way that it would yield `capacity` elements.
There is no benefit from making all operations on binary data read
in small, async steps directly from the file.
Since we know file sizes beforehand, it's less complex to read
the entire data blob into memory once, then operate on that
synchronously.