Optimize asymmetric 2-bit zero-point search

[Qubitium]

Summary

• evaluate both adjacent integer zero-point orientations during asymmetric 2-bit MSE range search
• select the scale and zero-point pair with the lower configured reconstruction objective
• add a synthetic regression covering a distribution where the alternate orientation reduces error

Root cause

An affine 2-bit grid has four codes. With zero point 2 its reconstructed levels are proportional to [-2, -1, 0, 1]; with zero point 1 they are proportional to [-1, 0, 1, 2].

The existing range search rounded the ideal affine zero point before measuring reconstruction error. At two bits, that rounding decision chooses which side receives the fourth level, so range geometry could select the orientation independently of the MSE objective.

For each scale candidate, this change retains the rounded zero point and also evaluates the other adjacent legal integer zero point. The previous candidate remains in the search set, so the selected local objective cannot become worse.

Scope

The additional candidate is restricted to bits=2, sym=False, non-groupwise affine quantization when MSE range search is enabled. Symmetric quantization, other bit widths, groupwise signed quantization, configuration, packing, and serialized formats are unchanged.

Validation

• pytest -q tests/test_gptq.py: 3 passed, 4 accelerator-only tests skipped
• synthetic regression selects zero point 1 and has lower MSE than the prior zero-point-2 candidate restriction
• critical Ruff rules (E9,F63,F7,F82) passed on changed files
• Python compilation and git diff --check passed