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>

test/lang/test_srt_backend.py

@@ -30,7 +30,9 @@ class TestSRTBackend(unittest.TestCase):
 
     @classmethod
     def setUpClass(cls):
-        cls.backend = sgl.Runtime(model_path=DEFAULT_MODEL_NAME_FOR_TEST)
+        cls.backend = sgl.Runtime(
+            model_path=DEFAULT_MODEL_NAME_FOR_TEST, cuda_graph_max_bs=4
+        )
         sgl.set_default_backend(cls.backend)
 
     @classmethod