Add integration tests

[Kovbo]

Adding a script to run integration tests.
The script defines a list of notebooks for testing, and starts a SkyPilot cluster executing notebooks one by one.
No need to SSH into the cluster, just run scripts/launch-test-cluster.sh --no-pull and it will execute tests, print results in the console, and terminate the cluster.
Currently, we test the most popular notebooks from the examples folder.

• I had to use openpipe-art==0.4.7 in all examples because it could not run notebooks programatically with the old version. To keep it compatible with Google Colab, we probably need to release a new art version that supports vLLM 0.9.2, and pin Colab examples to the old vLLM version. Can we?

• I was constantly running into the RuntimeError: CUDA error: an illegal memory access was encountered. Turns out, it happens when you start different vLLM engines one by one. For example, in one notebook, we build a vLLM engine with the default configuration, but in the new run, we provide engine_args=enforce_eager=True. It does not fully clean the memory, and the new engine fails.
  Workeround: set model._internal_config to None during tests. Not ideal since we cannot test configs.

• It overrides some notebook variables (to run only one training step for faster execution) and changes the project name so logs are recorded under the “Tester” project on W&B.

• Added --no-pull args to both launch-test-cluster.sh and launch-cluster.sh to deploy the current version without reverting to the main branch.

• Pytest caught some silent type errors that were not visible during regular execution. Minor changes in src/art/trajectories.py and src/art/unsloth/train.py

• I’m still occasionally hitting the following error during test execution: RuntimeError: Sleep mode can only be used for one instance per process. Not sure why it’s happening. Can we disable sleep mode for tests?