FIX: keep Responses web search calls replayable and behind the answer

[Utkarsh Bahuguna (u7k4rs6)]

Fixes #3079.

• Parse: store the whole web_search_call item (id, status, action) instead of just type and id.
• Replay: send exactly the fields the input item takes. Older rows that only have type and id can't be made valid, so they're skipped with a warning instead of failing the whole request.
• Order: tool records go after the text, same as reasoning, so get_value() is the answer again. Replay order doesn't change since top-level items are always emitted before the message item.

Same repro as the issue on this branch:

get_value(): PyRIT 0.13 is out.
replayed: {'id': 'ws_1', 'type': 'web_search_call', 'status': 'completed', 'action': {'type': 'search', 'query': 'pyrit release'}}

The old tests pinned the call_id/query shape, so I moved them to the real one. New tests: the answer is the response value, the replayed item matches the SDK param fields, and a legacy row gets skipped. They fail on main. tests/unit/prompt_target, backend, executor and score: 8853 passed (4 MCP/tool files didn't collect locally because httpx2 isn't installed; they don't touch this).

I didn't rerun the notebook since it needs the OpenAI endpoint, so its saved output still shows the old order.