fix(models): handle string content when merging consecutive same-role messages

[VANDRANKI]

Summary

get_clean_message_list raises AssertionError when the input contains two or more consecutive messages with the same role (e.g. two system messages) and those messages have plain string content.

Reproducer from issue #1972:

messages = [
    {"role": "system", "content": "Start every reply with FOO"},
    {"role": "system", "content": "End every reply with BAR"},
    {"role": "user", "content": "Say hello"},
]
LiteLLMModel(...)(messages)  # AssertionError

Root cause

ChatMessage.content is str | list[dict] | None. When messages are passed as plain dicts with a string "content" field, from_dict stores content as a str. The merge branch (hit when two consecutive messages share the same role) immediately asserts isinstance(message.content, list) without handling the string case, causing the crash.

Fix

Before attempting the merge, normalize both the incoming message.content and the previously-stored output_message_list[-1]["content"] to [{"type": "text", "text": ...}] if either is a string. The existing list-based merge logic then handles all cases uniformly, and consecutive string-content system messages are concatenated with a newline.

No behavior change for callers that already pass list-format content.

Fixes #1972