fix: default SLM_MODEL_NAME to openai/gpt-oss-20b for NVIDIA Build API

refactor: eliminate Qwen2.5-0.5B-Instruct local GPU loading, use pure NVIDIA Build API or 0-GPU fallback

feat: integrate NVIDIA Build API (openai/gpt-oss-20b) into SLM explanation layer with reasoning support

fix: resolve ZeroGPU CUDA init boundary in model_loader.py

feat: complete RazorShield AI risk platform