SGLang adds NVIDIA NVFP4 KV cache, 1M-context decode up to 78%

SGLang with Qwen and NVIDIA landed NVFP4 KV cache quantization, cutting KV memory to about 56% of FP8. Decode throughput rose 26-30% at equal concurrency and up to 78% at 1M context; agent workloads gain via higher cache hit rates.