Benchmarking Self-Hosted Gemma 2 9B vs. Frontier APIs: The FP8 Quantization Prefill Tax and VRAM Realities on an NVIDIA L4 [P] — PLINKFEED