Add conversational image editing and attachment improvements

[artcc]

Summary

Related to #135.

• Add edit_image so a tool-capable chat model can edit a selected image from the conversation using the configured dedicated image specialist and /images/edits.
• Add attachment thumbnails and previews, ordered multi-image selection from Photos, and an Attach Again action on conversation images.
• Scope asynchronous attachment and drag-and-drop preparation to the active conversation, preserving selection order and discarding stale results after navigation, cancellation, or a data reset.
• Share one durable image-operation attempt per user turn across generation and editing, including cancellation before transcript delivery, regeneration, legacy transcripts, backups, and branching.
• Serialize historical assistant images as user-side visual context for OpenAI-compatible Chat Completions, preserving tool-call/result ordering and the stored conversation.
• Add focused regression coverage and update specifications, localized strings, the changelog, and English/Spanish TestFlight notes.

How editing works

With a compatible dedicated image model selected under Image and Vision → Image Generation, the chat can resolve an earlier image by its attachment UUID and send the actual image plus the requested change to /images/edits. For example: generate a cat, then ask to change the background while keeping the same cat. Reattachment is optional when referring to an image already in the conversation; Attach Again lets the user select one explicitly.

Editing targets one image per request. The principal must support tool calling and lack native image generation; the selected dedicated image specialist must advertise vision and its provider must support image editing. Failed edits do not fall back to text-only generation.

Verification

• The maintainer manually tested generation followed by an edit of the same image and confirmed the flow worked.
• During implementation, the iOS simulator app and test products compiled successfully.
• The macOS app compiled successfully; only the expected entitlement warnings from disabling code signing remained.
• git diff --check passed during implementation.
• Regression tests cover request serialization, tool availability and shared attempt limits, cancellation/restoration, attachment lifecycle, drag-and-drop context changes, reattachment, and backup references. The final test sources were compiled but not executed.

Scope

This PR focuses on single-image conversational editing and attachment UX. Clipboard image paste, multi-image reference editing, and delegated editing through chat-based image specialists are outside this change.