`Sequence`s are covariant, so this allows you to pass in a
`list[np.ndarray]` without issues. From briefly looking at the
implementation, we do *not* modify the original array, so this is ok.

See https://mypy.readthedocs.io/en/stable/common_issues.html#invariance-vs-covariance
for more information.

These functions take `str`s, *not* `Format` objects. I'm not
sure what the intention of the `Format` objects were in the first place.

Without this, if you do `with imageio.get_writer(...) as writer`,
`writer` will have the type `Format._BaseReaderWriter` since
`__enter__` does not preserve *which* subclass we are.

typing.IO[bytes] is more general (for instance, `io.BytesIO` is an
`IO[bytes]` but not a `BinaryIO`), and we don't need any of the
functionality in `BinaryIO`.

Ideally, we would actually define a custom protocol here (since we
probably don't need the entire file-like API), but that would be a
lot of work.

The `@overload` still needs a *final* declaration -- otherwise,
some type-checkers will silently ignore the last overload.