Chunked Prefill Fixes LLM Server Freezes from Long Prompts

A single long prompt can freeze an LLM server because prefill (compute-bound) blocks decode (memory-bound) in naive schedulers. Chunked prefill splits prompts into fixed-size chunks interleaved with decode tokens, smoothing inter-token latency. The trade-off is time-to-first-token vs throughput, tunable via vLLM's max_num_batched_tokens parameter.