fix(proxy): harden image generation and edit requests

[buluw]

Summary

Make OpenAI-compatible image generation and edit requests preserve model, ordering, sizing, and input-boundary semantics.

Closes #3349.

Changes

• Resolve dynamic image aliases against the selected account for both generation and edit requests.
• Accept ordered Canvas-style data-URL references on image generations without fetching remote URLs.
• Preserve repeated image, image[], and image<digits> multipart fields in arrival order.
• Place an edit mask immediately after the first input image.
• Preserve model-suffix aspect ratios when no explicit ratio or recognized size is supplied.
• Enforce image_size, quality, and model-suffix priority consistently.
• Reject invalid image data and enforce 16 images, 20 MiB per decoded image, and 32 MiB total decoded input.

Validation

• cargo test --manifest-path src-tauri/Cargo.toml --offline --lib stream_peek_tests
  • 15 passed, 0 failed; six tests directly cover image generation/edit compatibility.
• cargo test --manifest-path src-tauri/Cargo.toml --offline --lib proxy::mappers::common_utils::tests
  • 13 passed, 0 failed.

Compatibility

• Text-only image generation remains unchanged.
• Existing routes and response formats remain unchanged.
• Invalid or oversized inputs now return HTTP 400 before an upstream request.
• Strict multipart MIME/empty-file rejection and an upper bound for n are not included.
• Real multipart and Canvas client integration remain to be verified separately.