CI nightlies cpu/gpu & cleanup

[aliberts]

This adds the ability to test lerobot on cpu & gpu docker builds

Changes:

• Add nightly cpu & gpu docker builds and testing
• PRs will be tested with pip install rather than poetry. This is to remove the necessity to have and maintain 2 different versions of pyproject.toml and poetry.lock.
  This does increase install time as poetry parallelizes package installs and is much faster (~30s vs ~3mn) but dependencies will be cached so it won't matter much in most cases where the PR doesn't change the pyproject.toml file.
• For now though, we still need another version of the pyproject.toml files for the docker builds: I need to install the gym envs as local path dependency after checking them out with ssh in the CI, rather than as direct git dependencies during the docker build which would require setting up ssh in the dockerfile. Since I do not want to go down that rabbit hole, I've set up a script to automatically change those git dependancies as local path dependencies. When the environments will be pip-installable (at release), all this hurdle will disappear.
• Add a DEVICE flag that shows up when running tests.
• Move end-to-end test inside a Makefile to just call make test-ete rather than having duplicates in the different workflows.

CI structure:

.github/workflows/
├── build-docker-images.yml     -> Builds gpu/cpu docker images every night from main
├── nightly-tests.yml           -> Runs tests on previously built docker images every night
├── test-poetry_DEPRECATED.yml  -> Former test.yml, soon to be removed
└── test.yml                    -> Tests (cpu) run on PRs and merge to main

Todo in future PRs (at release):

• remove the env repos checkout and local install
• change envs from git dependencies to just pip dependencies

[Cadene]

LGTM! Thanks ;)

I would suggest renaming ete -> end-to-end to be more explicit, but not sure.