This is apropos of the discussion in BLAKE3-team/BLAKE3#205.
We should try parallelizing single-threaded reads over multiple files. Single-threaded reads which use update_reader would allow for more granular progress bars, and might wind up being more efficient overall.