Migrate OpenAI to Responses API + reasoning + images

[datlechin]

Summary

• Closes Migrate OpenAI provider to Responses API and add Codex models #1112. .openAI provider now uses POST /v1/responses via new OpenAIResponsesProvider. OpenRouter/Ollama/Custom keep Chat Completions via existing OpenAICompatibleProvider.
• Adds reasoning panel (collapsible DisclosureGroup, "Reasoning…" while streaming, 80ms debounce on delta text). Surfaced for OpenAI reasoning models (gpt-5.5, gpt-5-codex, gpt-5.3-codex, gpt-5.4-mini) and Claude extended thinking (Opus 4.7, Sonnet 4.6, Haiku 4.5).
• Adds image input: drag-and-drop, paste, attachment chip. HEIC/TIFF/BMP convert to PNG/JPEG q0.92 via ImageIO. EXIF and GPS stripped at the boundary. Stored in ~/Library/Caches/com.TablePro/AIChatImages/<uuid>.<ext>; wire holds filename + media type only.
• Reasoning effort picker per provider, gated by descriptor.supportsReasoning and the model's supportedEffortLevels. OpenAI: minimal/low/medium/high/xhigh. Claude: low/medium/high/xhigh mapped to adaptive effort or budget_tokens on pre-adaptive models. max_output_tokens auto-scales per effort level so budget_tokens < max_tokens holds for Anthropic.
• Strict tools by default: ChatToolSchemaBuilder.object emits additionalProperties: false, auto-includes all property keys in required, supports nullable unions for optional fields. All 8 existing tools migrated.
• Stateless full-history replay with reasoning items round-tripped via ReasoningOpaque(itemID, value, blockType) — server-issued rs_* ID flows back into the next input array, no synthetic ID collisions. Required for Responses tool-loop correctness (server rejects function_call_output without its preceding reasoning item).
• Curated model picker (Picker.menu) with Curated / Fetched / Other sections plus free-text fallback.
• Zero PluginKit ABI impact (AI layer is app-only).

Test plan

• Build clean (xcodebuild ... build)
• SwiftLint clean (swiftlint lint --strict)
• 27 new tests pass (parser, encoding, strict tool audit) plus all existing AI tests
• Smoke test: OpenAI with gpt-5.5 + reasoning medium → Reasoning panel expands while streaming, collapses on done, encrypted reasoning item round-trips through tool loops without 400 error
• Smoke test: Claude Opus 4.7 with reasoning high → adaptive thinking, signature round-trips through tool loop
• Smoke test: drag PNG/HEIC into composer → chip appears → submit → image renders in assistant view → image block round-trips in ChatTurn history
• Smoke test: paste image from clipboard → same flow as drop
• Smoke test: change to Ollama provider → drop zone hidden, composer accepts no images
• Smoke test: execute_query tool still parses correctly after strict schema migration (optional fields sent as null or omitted)