Model Offloading Support Pt 2

[Satrat]

SUMMARY:
Adding support for partially offloading models to cpu. Requires this PR to compressed-tensors as well: neuralmagic/compressed-tensors#113. A couple main changes here:

• Similar to in the compressed-tensors PR, we need to be careful to update the offloaded state dict when changing any parameters
• helper functions for determining memory allocations, taking quantization/GPTQ into account or providing custom requirements.
• Removing the requirement that calibration data be passed for oneshot. Scale/zp for weights are set when calibration is initialized
• Adding examples for offloading oneshot runs

TEST PLAN:
Only manual testing so far, need to add in unit and integration tests after this lands