Add Hailuo H3 video generation support

[kapelame]

Summary

• route MiniMax-H3 video generation and task queries through the V2 API
• build H3 requests with typed text, image, video, and audio content while preserving the existing Hailuo 2.3 V1 path
• keep the CLI surface minimal with --image shared across legacy and H3 models, plus H3 reference media options
• validate H3 media roles, combinations, counts, decoded sizes, duration, ratio, and request size before sending
• support V2 polling and direct result downloads in the CLI and SDK
• document the new behavior in English and Chinese

Why

Hailuo H3 uses /v2/video_generation and a multimodal content[] request schema, which is incompatible with the fixed-field V1 request used by Hailuo 2.3. This change adds the V2 path without changing the existing 2.3 request behavior.

Local MP3 references are also encoded as the H3-compatible audio/mp3 data URI MIME type.

Validation

• bun run typecheck
• bun run lint (0 errors; 2 existing no-explicit-any warnings)
• bun test (434 passed, 0 failed)
• bun run build
• live Pay-as-you-go H3 generation:
  • image-to-video
  • reference-video generation
  • reference image + video + audio generation
• region auto-detection verified without passing --region

---

^{Need help on this PR? Tag @codesmith-bot with what you need. Autofix is disabled.}