fix(openai): deduplicate Copilot response text

[Sayt-0]

Summary

• deduplicate GitHub Copilot Responses API text by stable output_index when event item_id values differ
• preserve the completed-item snapshot fallback when no text delta was received
• add regression coverage for duplicated structured output, missing deltas, and multiple output slots

Issue expectations

Expectation	Implementation
Responses from github-copilot are emitted once	Streamed content is tracked by both item ID and output_index
Judge output remains valid JSON	The completed item snapshot is not appended after its text deltas
Providers that omit text deltas still return content	response.output_item.done remains a fallback when no content was emitted
Multiple outputs remain independent	Content state is tracked separately for each output_index

Reproduction

Before the fix, the regression stream produced:

{"answer":"structured","confidence":0.9}{"answer":"structured","confidence":0.9}

After the fix, it produces:

{"answer":"structured","confidence":0.9}

Validation

• go test ./pkg/model/provider/openai -run TestResponseStream_CopilotUnstableItemIDsOutputTextNotDuplicated -count=20
• go test -race ./pkg/model/provider/openai -run TestResponseStream_CopilotUnstableItemIDsOutputTextNotDuplicated -count=1
• go test ./pkg/model/provider/openai -count=1
• env -u HTTP_PROXY -u HTTPS_PROXY -u ALL_PROXY -u http_proxy -u https_proxy -u all_proxy task test
• task --force lint
• task build

Live verification against the affected gpt-5.6-* models was unavailable because the models are disabled for the available Copilot token and return model_not_supported. The regression test replays the observed Copilot event shape deterministically.

Fixes #3839