Add MiniMax image-to-image backend for garment and modeled stages

[octo-patch]

Reason: Add a MiniMax image-to-image backend so garment and modeled edits can run through the /v1/image_generation endpoint instead of only the OpenAI edits schema.

What changed

The import pipeline generated the garment and modeled stages exclusively through the
OpenAI multipart /images/edits schema and parsed data[0].b64_json. This adds an
alternative MiniMax image-to-image backend behind a provider switch, so the generated
bytes flow into the existing chroma-cleanup and review pipeline unchanged.

• miniMaxEdit() posts JSON to <base>/image_generation with Authorization: Bearer,
  mapping the one or two reference images to subject_reference
  ({ type: "character", image_file: "data:image/png;base64,..." }).
• Requests carry the supported fields: model (image-01 / image-01-live), prompt,
  response_format, width/height derived from the stage size (validated to the
  [512, 2048], multiple-of-8 range) with aspect_ratio taking priority when set, plus
  optional prompt_optimizer and seed. The prompt is clamped to the documented
  1500-character limit.
• Both response shapes are handled: data.image_base64[0] is decoded directly, and
  data.image_urls[0] is downloaded, returning PNG bytes either way. Non-zero
  base_resp.status_code surfaces status_msg as an error.
• A small editImage() dispatcher selects the backend from WARDROBE_IMAGE_PROVIDER
  (default openai), preserving the existing OpenAI behavior. Clothing detection still
  uses the OpenAI vision model.
• Global endpoint defaults to https://api.minimax.io/v1; the China endpoint
  (https://api.minimaxi.com/v1) is configurable via MINIMAX_API_BASE_URL.
• .env.example and the README configuration table document the new variables.

Checks

• npm run check (Vite production build) passes.
• Exercised miniMaxEdit() with a stubbed fetch: verified endpoint/auth/JSON headers,
  subject_reference mapping for one and two references, width/height vs
  aspect_ratio selection, prompt clamping, base64 and URL response decoding to PNG
  bytes, and error propagation on a non-zero status code.

[octo-patch]

Updated the MiniMax image response handling to decode base64 from data.image_urls and download URL responses, added n: 1 and seed configuration, and added focused tests. I ran npm test and npm run check.