[Tests] Use proper offloading utils in test_compress_tensor_utils

[kylesayrs]

Purpose

• Prerequisite for Use model compression pathways #1419
  • This PR disables getting the offloaded state dict unless necessary (sparsity statistics). However, the utility function cpu_offload only works if the offloaded state dict is retrieved. Let's replace this with dispatch_model, which is the actual function used by PretrainedModel, not cpu_offload

Changes

• Rename device_map to device
• Use dispatch_model rather than cpu_offload
• Use align_module_device and update_offload_parameter utilities
  • This change is necessary because, after these changes, some of these test models no longer have offloaded state dicts (which is the way it should always have been)

[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.

[brian-dellabetta]

💚