Fix CPU Offloading

[dsikka]

SUMMARY:

• When updating the training args, place_model_on_device was missed and as a result, when creating the trainer (which we really should not be doing during oneshot...) the default value is left as True and the trainer tries to move the model to a gpu, if it is available.
• We want this argument to be False as we handle the device map and model initialization based on the calibration needs

TEST PLAN:

• cpu_offloading_fp8.py ran to completion without issue
• mult_gpus_int8_device_map made it past the error and is running

[github-actions]

👋 Hi! Thank you for contributing to llm-compressor. Please add the ready label when the PR is ready for review.

Note: This is required to complete the testing suite, please only add the label once the PR is code complete and local testing has been performed.