There are two main things here:
 - `Box.flatten_n` and `Box.unflatten_n` should work on `Iterables`, so
   we need to use `len()` instead of checking the shape field.
 - Weights passed to `Discrete.weighted_sample` may not be normalized.
   Normalize them.

Also, add tests to make sure that these specific breakages don't happen
again.

Note that there is still a garage test failure due to incompatibilities
with the `gym.Box` constructor, but I'll fix that in garage.