enable native MPS GPTQ quantization

[Qubitium]

What changed

• stop models.auto from globally forcing PYTORCH_ENABLE_MPS_FALLBACK=1 on macOS
• remove obsolete fallback requirements from GPTQ, GPTAQ, and FOEM quantization
• add macOS regression coverage for environment preservation and native MPS GPTQ quantization

Why

Modern PyTorch MPS implements the GPTQ Hessian and quantization operations used here. The old guard forced unsupported operations onto CPU even when the Metal implementation was available, preventing native MPS execution and making Hessian math substantially slower on Apple Silicon.

Fallback remains available as an explicit PyTorch opt-in for users running older builds with unsupported operations.

Validation

• tests/test_mps_gptq.py: 2 passed on Apple M4 Max with PYTORCH_ENABLE_MPS_FALLBACK unset
• complete GPTQ quantization produced finite MPS weights without CPU fallback
• focused GPTQ configuration tests: 2 passed
• git diff --check
• Ruff passes for the new regression test