Quantization: lm_head quantization, k_quant_down heuristic, No zero offset while packing

[jambayk]

Describe your changes

• Gptq and olive quant checkpoint now supports lm_head quantization
• SelectiveMixedPrecision has a new heuristic called k_quant_down which is similar to k_quant_mixed but only takes the down_proj linears and lm_head.
• QuantLinear packing now doesn't do the gptq style -1 offset to the zero points while packing. This causes issues with asymmetric quantization when any zero point value is 0 as the values are expected to be unsigned int. It also now has a from_tensor method that can be used to create a new instance from already quantized weights.

Checklist before requesting a review

• Add unit tests for this change.
• Make sure all tests can pass.
• Update documents if necessary.
• Lint and apply fixes to your code by running lintrunner -a
• Is this a user-facing change? If yes, give a description of this change to be included in the release notes.
• Is this PR including examples changes? If yes, please remember to update example documentation in a follow-up PR.

(Optional) Issue link