fix(beta): persist OpenAI server-side builtin tool calls in history

[vvlrff]

Problem

The OpenAI Responses API requires that reasoning items be paired with their server-side tool calls (web_search, code_interpreter, image_generation) when history is replayed back to the model. Otherwise the API rejects the request.

Current behaviour was broken for replay:

• Server-side calls were emitted as generic BuiltinToolCallEvent / BuiltinToolResultEvent — with no reference to the original SDK object. When rebuilding history for a follow-up request there was nothing to re-emit.
• Reasoning items were emitted as ModelReasoning(text) — also with no SDK object attached.
• In the streaming path, some builtin items were processed on ResponseOutputItemAddedEvent, when the SDK object was not yet fully populated (e.g. outputs missing on code_interpreter).

Net effect: a conversation that used any builtin server-side tool could not be fed back to the Responses API in a subsequent turn.

Solution

New provider-specific events

File: autogen/beta/config/openai/events.py

Event	Base class	Payload
OpenAIServerToolCallEvent	BuiltinToolCallEvent	item: ResponseFunctionWebSearch | ResponseCodeInterpreterToolCall | ImageGenerationCall
OpenAIServerToolResultEvent	BuiltinToolResultEvent	observability-only, no payload
OpenAIReasoningEvent	ModelReasoning	item: ResponseReasoningItem

Each event carries the original SDK object (item) so the mapper can replay it verbatim.

History replay

File: autogen/beta/config/openai/mappers.py

New branch in events_to_responses_input:

elif isinstance(message, (OpenAIReasoningEvent, OpenAIServerToolCallEvent)):
    result.append(message.item.model_dump(exclude_none=True, mode="json"))

OpenAIServerToolResultEvent is intentionally not re-emitted — the Responses API models a server-side tool as a single combined item, already covered by the paired call event.

Client

File: autogen/beta/config/openai/openai_responses_client.py

_process_response (non-streaming):

• web_search / code_interpreter / image_generation now emit OpenAIServerToolCallEvent(item=...) followed by an empty OpenAIServerToolResultEvent.
• Reasoning is emitted as OpenAIReasoningEvent(text, item=...), where text joins all summary[].text with \n\n.
• Duplicate data (status / container_id / outputs) removed from ToolResult payloads — it all lives on item now.

_process_stream (streaming):

• All builtin / reasoning handling moved from ResponseOutputItemAddedEvent to ResponseOutputItemDoneEvent, so the SDK object is fully populated (on Added, code_interpreter.outputs is empty).
• Dropped provider_data on BuiltinToolCallEvent — this info is on item now.
• Removed the now-unused ResponseOutputItemAddedEvent import.

[vvlrff]

This pr can be considered after #2695