Add PyTorch-native K-quant pass

[jambayk]

Describe your changes

Add KQuant pass under olive.passes.pytorch.kquant that implements the ggml K-quant weight-only search in PyTorch. A unified _kquant_search driver covers both variants:

• Asymmetric (analogue of ggml's make_qkx2_quants used by Q2_K/Q4_K/Q5_K): tracks per-group (min, max), refines (scale, offset) via 2-D LSQ over a sweep of perturbed iscale factors.
• Symmetric (analogue of ggml's make_qx_quants used by Q3_K/Q6_K): uses max|x| as normalizer, zero point fixed at midq, refines scale via 1-D LSQ sumlx/suml2.

Outputs (scales, zero_points) shaped to match WeightQuantizer.find_qparams, with the search parameterized by (maxq, minq) pulled directly from each module's WeightQuantizer (via get_maxq_minq) so the algorithm and finalize agree on the integer range. Uses the shared prepare_model/finalize plumbing so embeddings and lm_head can be quantized and retied like the other PyTorch quant passes (RTN, GPTQ).

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.

Release note: New KQuant PyTorch pass for weight-only K-quant quantization (asymmetric + symmetric, 2/4/8-bit). Rtn and KQuant now also advertise uint2/int2 precisions.

(Optional) Issue link

[Copilot]

Pull request overview

Adds a new PyTorch-native KQuant pass for Hugging Face/PyTorch model weight-only quantization, using shared Olive quantization preparation/finalization flow and extending pass metadata to advertise 2-bit precisions.

Changes:

• Introduces olive.passes.pytorch.kquant with K-quant qparam search and pass implementation.
• Registers KQuant in olive_config.json and adds 2-bit precision support to Rtn.
• Adds PyTorch unit tests for KQuant qparam quality, pass execution, overrides, embeddings, and lm_head composition.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 3 comments.

File	Description
olive/passes/pytorch/kquant.py	Implements KQuant search, qparam generation, config validation, and pass execution.
olive/olive_config.json	Registers KQuant and updates Rtn supported precisions.
test/passes/pytorch/test_kquant.py	Adds coverage for KQuant qparams and end-to-end pass behavior.