Bsuite

Fixes a version discrepancy between PyPI and GitHub.

- Fixed a bug where some experiments were assigned to the wrong tag (736b9fe0fcdf42a7ccb887fe8727f781b590ff77).
- Add GitHub CI (pytype + pytest) and improve agent tests.
- Some minor quality-of-life changes for getting the bsuite_num_episodes constant.

- Fixed an issue with cartpole_swingup's `best_episode` not being computed correctly due to negative rewards (beb163021f56eb065508441d9e4d859580c409dc). This change will affect scores for this experiment, so we bump the minor version in keeping with semantic versioning.
- Fix versioning issues in colab and setup.py.
- Improve test.sh coverage.
- Make TAGS accessible as sweep (e.g. sweep.TAGS['exploration'] to retrieve all exploration experiments.
- Add a bootstrapped DQN baseline in JAX.
- Extract environments to their own package for easier direct access for debugging/development.

Contents include:

- Drop Python 2 support.
- Improve installation instructions.
- Upgrade baseline agents to TensorFlow 2 / JAX.
- Improvements to logging behaviour.