fix: enhance tool-call handling and tokenization in OpenAICompatibleTinkerServer

- Updated regex to correctly match tool-call tags, allowing for optional leading newline.
- Ensured content is set to None if no message content is present.
- Added byte encoding for tokens during tokenization for improved processing.

Author: bradhilton

src/art/tinker/server.py

@@ -134,14 +134,15 @@ async def chat_completions(
                             # from the content, so we remove them here.
                             content=(
                                 re.sub(
-                                    r"(?:\n<tool_call>.*?</tool_call>)+\s*$",
+                                    r"(?:\n?<tool_call>.*?</tool_call>)+\s*$",
                                     "",
                                     message["content"],
                                     flags=re.DOTALL,
                                 )
                                 if message["content"]
                                 else message["content"]
-                            ),
+                            )
+                            or None,
                             role="assistant",
                             tool_calls=[
                                 ChatCompletionMessageFunctionToolCall(
@@ -160,6 +161,9 @@ async def chat_completions(
                             content=[
                                 ChatCompletionTokenLogprob(
                                     token=f"token_id:{token}",
+                                    bytes=list(
+                                        renderer.tokenizer.decode(token).encode()
+                                    ),
                                     logprob=logprob,
                                     top_logprobs=[],
                                 )