Fix: Support Dict-Style Messages in get_clean_message_list

Summary

This pull request updates the get_clean_message_list() method in models.py to support dictionary-style messages in addition to structured Message objects. This improves developer experience and aligns TransformersModel with InferenceClientModel, which already supports dict input formats.

Problem

Passing a dictionary-style message to TransformersModel raises an AttributeError, due to direct attribute access (message.role) instead of dictionary key access:

AttributeError: 'dict' object has no attribute 'role'

File ".../smolagents/models.py", line 291, in get_clean_message_list
   role = message.role
          ^^^^^^^^^^^^

This breaks many natural use cases like:
model([{"role": "user", "content": [{"type": "text", "text": "Hello"}]}])

What Was Changed

Replaced attribute-style access (message.role, message.content) with dictionary-style access (message['role'], message['content']) throughout the get_clean_message_list() function.

Ensured message merging, image processing, and content flattening still function correctly.

Prevented crashes for users providing standard dict inputs, improving ease of use and compatibility.

Also added a test case to validate this functionality.
Test passed locally with pytest -k test_transformers_model_with_dict_message

Happy to add documentation or additional test coverage if needed.