feat: Add product widget response

[dsowinski2]

This PR adds possibility to render a product widget.

1. Adds a direct return tool which can handle returning product widget response.
2. Adds new agent callback handler class responsible for streaming widgets
3. Adds callback handler to injector so tool can stream
4. Handle messages and stream on Chat UI
5. Adds new message type (widget)
6. Add new field on Message model (widget_data)
7. Saves widget response to conversation history

[rafalcymerys]

@dsowinski2 I'm missing some context when it comes to how this is supposed to work. Could you add a description to this PR that describes the overall concept behind widget responses, how they're generated and managed?

[dsowinski2]

I've updated description and resolved conflicts, let me know if you need more details.

[mateus-po]

Two minor suggestions to the PresentProductTool implementation, otherwise LGTM

[mateus-po]

Maybe it would be a good idea to prefetch all products first? By implementing a get_by_ids(ids: list[str]) method on a product registry that would fetch all products in one DB call. Not a big issue, but would speed up the tool + would reduce the number of DB calls.

[mateus-po]

I am not a big fan of these 3 if self._streaming: blocks in the run method. I would do it like this:

1. Define a private method like:

def _handle_streaming(step: 'start' | 'in_progress' | 'end') -> void:
    if not self._streaming: return
    # remaining logic based on the streaming step

2. use it inside of the run method

It would make the code inside of the run method more readible and the whole tool class more modular.

[mateus-po]

Or more something like this (so that you can pass the streaming message data):

def _handle_streaming(step: 'start' | 'in_progress' | 'end', data: Optional[str]) -> void:
    if not self._streaming: return
    # remaining logic based on the streaming step