feat(sudoku): implement environment

[Egiob]

This PR contains all the changes related to the addition of the Sudoku environment. See below the principal stuff that needs to be discussed before merging (in my opinion).

Generator

Currently the environment uses a default DatabaseGenerator that resets to a new puzzle included in a pre-loaded database. These data are included under jumanji.environments.logic.sudoku.data. For now two databases are included: 1000_very_easy_puzzles.npy containing 1000 puzzles with >= 46 clues and 10000_ mixed_puzzles.npy containing 10000 puzzle of various difficulties. The exact number of puzzles may be changed so that (1) the "very-easy" dataset is indeed very-easy for A2C and (2) the "mixed" dataset contains enough diversity while not taking too much memory in the repo.

Versions

Two envs have been registered: Sudoku-v0 using by default the 10000_mixed_puzzles.npy database and Sudoku-very-easy-v0 using by default the 1000_very_easy_puzzles.npy database. For simplicity purpose, we propose to train the A2C agent on this latter environment.

Remaining actions:

• Check that the training on the very-easy dataset meets expectations
• Update Sudoku GIF and Image accordingly
• Revert changes in notebooks

References:
This is the paper I used to define that >=46 clues was "very-easy".

[CLAassistant]

All committers have signed the CLA.

[clement-bonnet]

Thank you for your contribution. Overall, looks very good to me. I have a few comments and NIT suggestions.
We can merge once we have a training curve (working), I'm wondering if we can make the network equivariant to permutation of digits at least.
Please let me know if you have any questions.

[clement-bonnet]

Thanks for the very nice contribution. I left a couple of nit comments. All good to me otherwise.