Disable parallel quantization for mul1 codebook

[remichu-ai]

Summary

convert.py --parallel_mode is currently unsafe with --codebook mul1. In a full GLM-5.2 MUL1 quant run it corrupts layer-0 EXL3 scale metadata (.suh tensors containing NaN/Inf), which then poisons layer-1 calibration Hessians and causes Cholesky failure.

This PR disables parallel_mode automatically when the selected codebook is mul1, with a warning, while leaving the existing parallel path unchanged for the default MCG codebook.

Validation

• Syntax check: python -m py_compile exllamav3/conversion/convert_model.py
• Reproduced failure with --codebook mul1 --parallel_mode: layer 0 produced NaN/Inf .suh; layer 1 Hessians became all-NaN and Cholesky failed.
• Retested GLM-5.2 full 3.0bpw MUL1 quant with serial mode: cleanly passed layers 0, 1, and 2 and reached loading layer 3. Layer-0 saved tensors had no NaN/Inf floating tensors.

Notes

This is a conservative correctness guard. It avoids the known-bad combination without changing quantization kernels or the parallel path for other codebooks.