feat(models): add GLM-5 Next quantization support

[ZX-ModelCloud]

Summary

• make dequantized safetensors output resumable and enforce a compatible contiguous layout
• register glm5_next with hybrid KDA/DSA, dense and routed-expert quantization coverage
• preserve the auxiliary MTP checkpoint tensors while quantizing all loaded decoder layers
• preserve DSA replay indices and GLM expert clamp behavior during calibration
• add model-quality expectations and update the supported-model documentation
• require defuser>=0.0.27 from feat: support GLM-5 Next expert defusion Defuser#66

Evaluation

BF16 ARC Challenge baseline (1,172 samples):

• acc: 0.49146757679180886
• acc_norm: 0.5273037542662116

Tests

• GLM-5 Next support tests: 7 passed
• dequant tests: 10 passed; the optional bitsandbytes case was not run because that package is unavailable in the test environment
• GLM-5 Next model eval test collection: 1 test collected