doc: add bench_one_batch_server in the benchmark doc (sgl-project#8441)

Author: Qiaolin-Yu

docs/references/benchmark_and_profiling.md

@@ -4,10 +4,15 @@
 
 - Benchmark the latency of running a single static batch without a server. The arguments are the same as for `launch_server.py`.
   Note that this is a simplified test script without a dynamic batching server, so it may run out of memory for a batch size that a real server can handle. A real server truncates the prefill into several batches, while this simplified script does not.
+  - Without a server (do not need to launch a server)
+    ```bash
+    python -m sglang.bench_one_batch --model-path meta-llama/Meta-Llama-3.1-8B-Instruct --batch 32 --input-len 256 --output-len 32
+    ```
+  - With a server (please use `sglang.launch_server` to launch a server first and run the following command.)
+    ```bash
+    python -m sglang.bench_one_batch_server --base-url http://127.0.0.1:30000 --model-path meta-llama/Meta-Llama-3.1-8B-Instruct --batch-size 32 --input-len 256 --output-len 32
+    ```
 
-  ```bash
-  python -m sglang.bench_one_batch --model-path meta-llama/Meta-Llama-3.1-8B-Instruct --batch 32 --input-len 256 --output-len 32
-  ```
 
 - Benchmark offline processing. This script will start an offline engine and run the benchmark.