Fix integer overflow in MultiDiscrete.flatten()

[olipinski]

Description

Fixes integer overflow in the flatten function when dtype is small.

The current code will calculate the offsets as is, overflowing the max integer value and going to negatives. Once this is passed to one_hot we get an error as per the linked issue.

File "C:\Users\user\Anaconda3\envs\TRG\lib\site-packages\gym\spaces\utils.py", line 90, in _flatten_multidiscrete
    onehot = np.zeros((offsets[-1],), dtype=space.dtype)
ValueError: negative dimensions are not allowed

Fixes

#53

Type of change

• Bug fix (non-breaking change which fixes an issue)

Test

import gym
import numpy as np

space = gym.spaces.MultiDiscrete([101, 101, 101, 101], dtype=np.int8)
x = space.sample()

y = gym.spaces.flatten(space, x)

x = gym.spaces.unflatten(space, y)

print(x)

[82 52 50 66]

Checklist:

• I have run the pre-commit checks with pre-commit run --all-files (see CONTRIBUTING.md instructions to set it up)
• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes

[pseudo-rnd-thoughts]

Could you explain the error that occurred before running the example code

Could you add a test in tests/spaces/test_discrete.py using the example provided in the PR.

[olipinski]

Updated, and test added.

[pseudo-rnd-thoughts]

I think changing to int32 is good enough. It is impractical to anyone to have more subspaces than ~2 billion so I think this is a safe choice.