Rewrite README: minimal style, drop perf data, match nvidia M2.5 recipe details

Add model card: recipe, hardware support, vLLM launch config, benchmarks

NVFP4 quant via modelopt 0.43.0rc3 + TensorRT-Model-Optimizer examples, recipe general/ptq/nvfp4_experts_only-fp8_kv (calib_size=1024, moe_calib_experts_ratio=1.0, cnn_dailymail + Nemotron-Post-Training-Dataset-v2, batch_size=64->32 auto-fallback, source=FP8 M2.7, FP8->bf16 on-forward dequant, 4x H100 SXM)

initial commit