Add CI for XGrammar

[Ubospica]

XGrammar needs CI to ensure stability and help review community contributions. Guidelines for CI:

1. Currently CI only needs to run on CPU (tests requiring GPU can be skipped, specifically the GPU kernel tests)
2. Run linter (pre-commit hooks, ruff)
3. Test on multiple platforms (Linux, MacOS, Windows) x (x64, ARM64)
4. Test C++ and Python code (ctest and pytest)
5. CI should be enforced to every commit on main branch, every PR, and can be manually triggered
6. The model folder ~/.cache/huggingface should be shared between different runs
   Contribution is welcome!

[zbowling]

That new build_and_release.yaml should run all the CPU tests (excluding HF_TOKEN needed tests, but it could do those too if an HF_TOKEN is added to GitHub secrets, but these tests take significantly longer to run). It doesn't run C++ tests at the moment, though, but I can send PR after #190 lands to do that :)

[Ubospica]

@zbowling Yes, the build_and_release workflow already runs Python tests. But to make is more clear I think the C++ and Python tests can be separated, so build_and_release is only responsible for testing whether the package can be built, while run_tests is responsible for running all the tests. The second workflow relies on the first for the built package and Github artifacts for file transfer.

Regarding the HF_TOKEN needed tests, I have added a secret secrets.HF_TOKEN that has access to all models in the tests. If the model can be cached, I think these tests should be executed faster.