Convert environments to fixed horizon

[AdamGleave]

• For MuJoCo tasks, swap to using implementations from benchmark_environments. Not much "implementation" of ours to remove: just removing some registration statements, since we were just doing a thin wrapper around Gym.
• For PointMass,
• Fix bugs I found in the PointMass and Hopper ground-truth reward models. This is unlikely to meaningfully change results: regression and preferences were trained directly on the reward model anyway, and the differences were not the kind to change optimal policy so should make little difference to IRL.

[codecov[bot]]

Codecov Report

Merging #28 into master will decrease coverage by 0.01%. The diff coverage is 76.92%.

@@            Coverage Diff             @@
##           master      #28      +/-   ##
==========================================
- Coverage   87.25%   87.23%   -0.02%     
==========================================
  Files          54       54              
  Lines        3608     3595      -13     
==========================================
- Hits         3148     3136      -12     
+ Misses        460      459       -1     

Impacted Files	Coverage Δ
..._rewards/analysis/dissimilarity_heatmaps/config.py	66.23% <0.00%> (ø)
src/evaluating_rewards/experiments/env_rewards.py	0.00% <ø> (ø)
tests/test_rewards.py	100.00% <ø> (ø)
src/evaluating_rewards/envs/point_mass.py	83.90% <75.00%> (-0.10%)	:arrow_down:
src/evaluating_rewards/envs/__init__.py	100.00% <100.00%> (ø)
src/evaluating_rewards/envs/mujoco.py	98.16% <100.00%> (ø)
...ewards/analysis/dissimilarity_heatmaps/heatmaps.py	90.00% <0.00%> (+1.25%)	:arrow_up:

---

Continue to review full report at Codecov.

Legend - Click here to learn more Δ = absolute <relative> (impact), ø = not affected, ? = missing data Powered by Codecov. Last update 41bd8b8...c972465. Read the comment docs.

[AdamGleave]

Tests are failing because we're pointing at the benchmark-environments master which does not include my early registration. They pass on my local machine.

When https://github.com/HumanCompatibleAI/benchmark-environments/pull/12 gets merged, update requirements.txt to new SHA and re-run. It may make sense to wait for the other PRs to get merged too.