Add LongCat-Next integration and Azure endpoint judge support

[nssmd]

Summary

• add LongCat-Next and LongCat-Next Visual CoT model integrations to lmms_eval
• add Azure OpenAI compatible endpoint judge support backed by Azure CLI bearer tokens and /home/aiscuser/endpoint.json style endpoint lists
• simplify evaluation scripts to use the active Python environment directly
• add a test_longcat_next.sh helper with default 8 visible GPUs and batch size 4
• allow LongCat-Next wrapper to pass max_memory through to transformers loading

Testing

• python -m py_compile on new / modified Python files
• bash -n on updated shell scripts
• validated Azure OpenAI compatible endpoint connectivity using Azure CLI token + endpoint list