Forensic documentation of deploying Google Gemma 4 27B-A4B-it (MoE) on Vertex AI with vLLM — 20 failure modes, dependency conflicts, and a community call-to-action for Multi-LoRA, quantization, vision & thinking support