FIX: Prevent CUDA context initialization due to AWQ

[BenjaminBossan]

Importing from AWQ triggers CUDA context initialization, which can be problematic in some circumstances (see #1877). This PR moves the import so that it's local, preventing this issue.

To test this, run this script:

import multiprocessing

from torch import nn
import peft  # noqa F401

def func():
    nn.Linear(2, 3).cuda(0)

if __name__ == "__main__":
    proc = multiprocessing.Process(target=func)
    proc.start()
    proc.join()

Ideally, we can add this to our nightly GPU tests, but running this from the pytest runner does not work (IIRC) so some extra steps are required.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.