[Transform] Fix accelerate import to keep it as optional dependency

[tboerstad]

PR #445 introduced a direct import of has_offloaded_params from accelerate.utils at the module level in transform/apply.py.

This could lead to upstream projects, like SGLang, failing to run with:

  File ".../compressed_tensors/transform/apply.py", line 18, in <module>
    from accelerate.utils import has_offloaded_params
ModuleNotFoundError: No module named 'accelerate'

This PR just changes to has_offloaded_params from compressed_tensors.utils.offload instead of accelerate.utils directly. This ensures accelerate remains optional and the package can be imported without it installed.

The utils.offload module already has proper fallback handling with @check_accelerate decorator that returns False when accelerate is not available.

I have run basic tests checking that imports work, and also the tests/test_transform/. I'm new to this project, so please let me know if there's something else needed.

[tboerstad]

To the maintainers:

The .yaml workflows (except for quality-check.yaml) all install [dev,accelerate] iiuc, so this was not caught in CI.
You could consider running some tests/or just import compressed_tensors in a CI job which only does not install any extras to avoid this happening in the future.

I don't want to mess with the CI as I'm new

[kylesayrs]

Awesome catch, thank you!

[kylesayrs]

@tboerstad You're absolutely right. This is somewhat difficult to catch as of now because the CI usually installs accelerate. At the moment it's probably not worth running additional CI runs without accelerate, so I'm glad this was caught.