fix(docker): remove dangling hybrid_modelopt_ct import in glm52 image
The production-hybrid-1.4 base image patches
vllm/model_executor/layers/quantization/__init__.py to add
'from . import hybrid_modelopt_ct' but never ships hybrid_modelopt_ct.py.
This breaks ALL vllm imports with a circular-import error:
ImportError: cannot import name 'hybrid_modelopt_ct' from partially
initialized module 'vllm.model_executor.layers.quantization'
Root cause: the 1.4 image's __init__.py references a module that doesn't
exist anywhere in the image (confirmed: find / returns nothing). The 1.2
image (which works) only imports hybrid_mxfp4_ct. The quantization dir
contents are otherwise identical between 1.2 and 1.4.
Fix: sed-delete the dangling import line after FROM 1.4. No recipe uses
hybrid_modelopt_ct (all use hybrid_mxfp4_ct), so removal is safe. The
NF4 kernel (nf4_e8m0 in b12x_moe.py) that 1.4 adds over 1.2 is preserved.
Verified on dgx1: rebuilt :glm52 image imports vllm successfully
(0.11.2.dev279+eldritch.final), NF4 kernel present, running container
unaffected (uses old image ID).
