fix(v2): retain legacy function_call in dump_message when content is empty - #2484

Problem

dump_message (instructor/v2/core/messages.py) only appended the legacy OpenAI function_call when the assistant message had truthy content. Tool/function-call responses commonly use content=None, so the serialized message dropped the function call entirely.

Because dump_message builds the assistant turn on the retry/reask path (OpenAI and Mistral handlers), a re-ask could omit the exact structured output that caused the validation failure, leaving the next request without the relevant assistant response.

Fixes #2464.

Repro (before)

from openai.types.chat import ChatCompletionMessage
from openai.types.chat.chat_completion_message import FunctionCall
from instructor.v2.core.messages import dump_message

msg = ChatCompletionMessage(role="assistant", content=None,
                            function_call=FunctionCall(name="lookup", arguments='{"id":7}'))
print(dump_message(msg))
# {'role': 'assistant', 'content': ''}   # function_call lost

Fix

Remove the falsy-content guard so a present function_call is always serialized. Behavior for non-empty string content and for list content (text/refusal flattening) is unchanged.

Tests

Added test_dump_message_retains_function_call_when_content_is_empty, which fails on main and passes with this change. Existing dump_message tests and the surrounding message/handler suites remain green. CHANGELOG updated under [Unreleased].