openai[patch]: fix get_num_tokens for function calls (#25785)

Closes https://github.com/langchain-ai/langchain/issues/25784

See additional discussion
[here](0a4ee864e9 (r145147380)).

libs/partners/openai/langchain_openai/chat_models/base.py

@@ -947,7 +947,7 @@ class BaseChatOpenAI(BaseChatModel):
                 else:
                     # Cast str(value) in case the message value is not a string
                     # This occurs with function messages
-                    num_tokens += len(encoding.encode(value))
+                    num_tokens += len(encoding.encode(str(value)))
                 if key == "name":
                     num_tokens += tokens_per_name
         # every reply is primed with <im_start>assistant

libs/partners/openai/tests/unit_tests/chat_models/test_base.py

@@ -677,7 +677,10 @@ def test_get_num_tokens_from_messages() -> None:
         AIMessage(
             "",
             additional_kwargs={
-                "function_call": json.dumps({"arguments": "old", "name": "fun"})
+                "function_call": {
+                    "arguments": json.dumps({"arg1": "arg1"}),
+                    "name": "fun",
+                }
             },
         ),
         AIMessage(
@@ -688,6 +691,6 @@ def test_get_num_tokens_from_messages() -> None:
         ),
         ToolMessage("foobar", tool_call_id="foo"),
     ]
-    expected = 170
+    expected = 176
     actual = llm.get_num_tokens_from_messages(messages)
     assert expected == actual