feat: support per-model Responses API protocol

[guima-why]

Summary

• Add per-model apiMode routing. Existing providers and Qwen models default to Chat Completions; the three registered OpenAI GPT-6 models default to Responses.
• Add dedicated OpenAI and standard DashScope Responses adapters for text/images, streaming, local function tools, usage, and stateless replay with store=false.
• Preserve native output items in local session history, with visible-message fallback after model/protocol/endpoint changes. Validate tool turns before execution and compact locally once on context-limit errors.
• Keep existing thinking configuration precedence and fall back from unsupported saved effort values using the selected Responses profile's defaults.

Validation

• Rebased onto cc7c9777 (current main). No textual conflicts; git range-diff confirms the feature patch is unchanged. Both branches' changed-file blobs and all six translation catalogs were preserved.
• Three focused compatibility reviewers found no P0/P1/P2 conflicts between the new AGUI structured resource-selection resume and Responses changes.
• After rebase: provider/Responses session/context/session storage/Web settings/AGUI/i18n regression: 1334 passed, 6 live tests skipped.
• After rebase: resource selector/A2A events/Agent tool batching regression: 629 passed.
• Python 3.10–3.14 focused feature regression: 479 passed per version (before rebase, identical feature code).
• make lint and git diff --check passed.

Real-model coverage and known limitations

• Live verification used DashScope Qwen only; no live GPT-6 verification.
• Seven real A2A scenarios were run with both Chat Completions and Responses: four scenarios passed both protocols, including backup-only restoration; three restart-recovery scenarios failed both protocols at the existing execution-control gate before resumed LLM dispatch.
• Those three recovery failures were independently reproduced with the unchanged pre-feature baseline. They remain limitations of the complete recovery regression.
• All cloud resources created by these runs were deleted. Local design and detailed regression artifacts are not included in this PR.