Using lazy sequences by default is too opinionated. They incur a lot of overhead and are tricky to manage.
Rather, we can expose a mechanism similar to `ocaml-h2` that uses a `schedule_read` function expecting two callbacks.

(for already closed stream)
anmonteiro/ocaml-h2#175

When multiple messages are batched and read into the buffer, only the
first message was extracted. Remaining messages would not be extracted
until the next invocation of `on_read`.