Support penalty in overlap mode; return logprob with chunked prefill; improve benchmark scripts (#3988)

Co-authored-by: SangBin Cho <rkooo567@gmail.com>
Co-authored-by: dhou-xai <dhou@x.ai>
Co-authored-by: Hanming Lu <hanming_lu@berkeley.edu>

python/sglang/lang/backend/runtime_endpoint.py

@@ -329,7 +329,12 @@ class RuntimeEndpoint(BaseBackend):
 
 def compute_normalized_prompt_logprobs(input_logprobs):
     values = [x[0] for x in input_logprobs if x[0]]
-    return sum(values) / len(values)
+    try:
+        return sum(values) / len(values)
+    except TypeError:
+        print(f"{input_logprobs=}", flush=True)
+        print(f"{input_logprobs[0]=}", flush=True)
+        exit(-1)
 
 
 class Runtime: