Jev2 vision — FP8 dynamic W8A8

Quantized from bqbbao6/b-jev with LLM Compressor FP8_DYNAMIC on Linear layers. lm_head stays in its original dtype because Jev derives scores from the logprob difference of two decision tokens. The Jev request schema and serving code are copied from the source export. Use the vLLM comparison script in quantize_jev to validate both text and image requests.