Add regression tests

[aliberts]

Review filter

What this does

• Add tests/scripts/save_policy_to_safetensor.py to generate test artifacts
• Add test_backward_compatibility to test generated outputs from the policies against artifacts:
  • output_dict
  • grad_stats (actually mean and std for each of the policy module's gradient)
  • param_stats (actually mean and std for each of the policy's module)
  • actions

Note: Due to floating point precision errors, I had to restrict this test to only be run on an x86_64 kernel, which is what the test artifacts were generated on. This shouldn't be an issue as our CI runs on this platform.

Thanks to @alexander-soare and @Cadene for your precious help on that one!

How was it tested?

I've set rtol=0.1 and atol=1e-7 for the assert torch.isclose() through trial & error (see failed tests on commits below)

How to checkout & try? (for the reviewer)

git checkout main
git fetch origin
git checkout -b user/aliberts/2024_04_29_add_regression_tests --track origin/user/aliberts/2024_04_29_add_regression_tests

DATA_DIR=tests/data pytest -v tests/test_policies.py::test_backward_compatibility

[alexander-soare]

Thanks a lot for this @aliberts. Some comments, mostly minor.

[Cadene]

YEASSSS PLEASE