Add code coverage, more end-to-end tests

[aliberts]

Note: I've split #54 into 2 PRs since they cover mainly unrelated changes

This adds:

• test_aloha in test_envs.py which was missing
• Code coverage
• More end-to-end tests:
  • train ACT on ALOHA (no online, offline_steps=2)
  • eval ACT on ALOHA (from train checkpoint)
  • eval ACT on ALOHA (no checkpoint, policy is None)
  • train Diffusion on PushT (no online, offline_steps=2)
  • eval Diffusion on PushT (from train checkpoint)
  • eval Diffusion on PushT (no checkpoint, policy is None)
  • train TDMPC on Simxarm (with balanced offline/online_steps=1
  • eval TDMPC on Simxarm (from train checkpoint)
  • eval TDMPC on Simxarm (no checkpoint, policy is None)

TODO

• Link repo to HF's Codecov account

[Cadene]

Could we have something in README in ### Contribute section indicating how to install dev dependencies? Are they installed when we do poetry install the first time?

[aliberts]

They are installed by default with poetry install. They can be skipped with poetry install --without dev.
Group dependencies are a conveniency offered by poetry for several purposes (one of which is staged build). In our case we just use it to differentiate those dev dependencies from the core ones (but it's really just labeling here)

[Cadene]

nice!

[Cadene]

The CI took 12mn to run. It's too much. What could we do to parallelize? Not mandatory, but it would be nice to address this in this PR.

[aliberts]

The CI took 12mn to run. It's too much. What could we do to parallelize? Not mandatory, but it would be nice to address this in this PR.

Once merged, dependencies will be cached so it should take ~4mns less.
Here are my thoughts on how to cut down time further:

• cache libegl1-mesa-dev install (~30s less)
• remove the ACT train on Aloha with no policy test, or at least reconfigure it (takes ~2mn right now, it shouldn't be that long)
• More generally change parameters for all end-to-end tests (I'm thinking about the number of steps, horizon, num_slices, batch_size...)

Do you see any other params that could impact run time significantly?

I will try those ideas before merging to see what works best.

[aliberts]

I'm merging this with the eval ACT on Aloha (policy is None) test deactivated so that we can move on for now.
I will take care of this in a further PR.