[GCP] Avoid dumping cachetools for backward compatibilty

[Michaelvll]

cachetools > 5.0 will have different APIs. Dumping the Cache will cause backward compatibility issues.

To reproduce:

1. pip install "cachetools<5.0"; sky launch -c test --cloud gcp --cpus 2
2. pip install "cachetools>=5.0"; sky launch -c test

Tested (run the relevant ones):

• Code formatting: bash format.sh
• Any manual or new tests for this PR (please specify below)
  • Reproducible code above
• All smoke tests: pytest tests/test_smoke.py
• Relevant individual smoke tests: pytest tests/test_smoke.py::test_fill_in_the_name
• Backward compatibility tests: bash tests/backward_comaptibility_tests.sh

[concretevitamin]

To reproduce:

1. pip install "cachetools<5.0"; sky launch -c test --cloud gcp --cpus 2
2. pip install "cachetools>=5.0"; sky launch -c test

How was this encountered? I feel like protecting against these kinds of out-of-band dependency breakage is hard, and we probably have other issues of this type. That said, ok to let this go in if important.

[Michaelvll]

To reproduce:

1. pip install "cachetools<5.0"; sky launch -c test --cloud gcp --cpus 2
2. pip install "cachetools>=5.0"; sky launch -c test

How was this encountered? I feel like protecting against these kinds of out-of-band dependency breakage is hard, and we probably have other issues of this type. That said, ok to let this go in if important.

This can happen when people upgrade their GCP package like google-auth which removes cachetools<5 constraint in their setup in a certain version. This issue is very serious, as after that all the GCP clusters will not be able to be loaded from our database, and a user may fail to find a way to recover it especially if they launched a new cluster with the new cachetools installed, i.e., both old and new version of clusters exist in the db.
I was only able to get myself out of the state by removing the new cluster from my database manually and downgrade my cachetools.

This kind of dependency package issue will only happens when we save the object with the package to the database with pickle.dump, and the package changes its API of object construction, so in the future we should be careful with what to save to the database, and avoid saving third-party objects.

[concretevitamin]

Thanks @Michaelvll, SGTM.