feat(animation-asset): add animation asset generation workflow

[adamhuang17]

Summary #28
Supports a single reference asset supplied either as an uploaded PNG, JPEG or as a public HTTP(S) URL, with upload-size, image-content, and request-parameter validation.
Builds a fixed six-phase animation prompt directly from the user-provided action_prompt, while instructing the model to preserve the subject’s identity, proportions, design, camera angle, and rendering style.
Makes exactly one image-editing request to QNAIGC using openai/gpt-image-2 and requires a 1536×1024, 3×2 Sprite Sheet. A deterministic local mock provider is also included for validating the complete workflow without consuming model quota.
Creates the Manifest before generation and deterministically splits the Sprite Sheet into six 512×512 PNG frames in row-major order, with either preserved positioning or optional bottom-center alignment.
Builds a six-frame GIF and performs structural quality checks for frame count, dimensions, empty content, Alpha coverage, edge contact, and exact duplicate frames.
Exposes background-job status and asset retrieval endpoints, and persists request metadata, the compiled generation prompt, Manifest, raw and processed frames, quality report, Sprite Sheet, and GIF.
Testing
Added automated tests covering:

• Multipart image uploads and public URL reference inputs.
• Direct inclusion of the user action description in the generated prompt.
• A single QNAIGC image-editing request and Base64 response parsing.
• Deterministic row-major splitting into six 512×512 frames.
• End-to-end service packaging with one provider call, six PNG frames, and a six-frame GIF.

Known limitations

• Each task currently supports only one reference asset; combined character, prop, or environment references are not supported.
• Output is fixed to six frames, a 3×2 layout, and 1536×1024; dynamic frame counts and adaptive layouts are not supported.
• Model output must exactly match the required dimensions. There is no automatic retry, resizing repair, or per-frame regeneration.
• Background removal or matting is not integrated, so generated backgrounds are preserved in the final frames.
• Motion continuity, subject consistency, and cell isolation still depend on the image model following the prompt.
• Job state is stored in memory and assets are written to the local filesystem. Job state does not survive service restarts and is not designed for distributed deployment.