fix(knowledge): support Doubao multimodal online embeddings (#1299)

[lymerin]

Summary

Fixes #1299.

Support Doubao vision embedding models in knowledge-base online embeddings. These models use /embeddings/multimodal, whose request and response formats differ from the OpenAI text embedding API.

• Use the multimodal protocol for doubao-embedding-vision-* models or an explicitly configured /embeddings/multimodal URL.
• Send typed text inputs and read the dense vector from data.embedding.
• Apply the same protocol selection to provider probing, document indexing, and retrieval.
• Request one vector per chunk to preserve chunk/vector alignment. Existing text embedding batching remains unchanged.

Endpoint IDs (ep-*) are supported through an explicitly configured multimodal URL.

Target branch

• Base is develop (feature / fix — default)
• Base is main (release/* or hotfix/* only)

Type of change

• Bug fix
• New feature
• Breaking change
• Documentation
• Refactor / chore
• Release / hotfix

Test plan

• make all passes locally — 3920 passed, 120 skipped
• Added/updated tests

Regression tests cover protocol selection, explicit endpoint URLs, endpoint IDs, request payloads, authentication/custom headers, ordered vectors across 21 chunks, and indexing/retrieval through the real SQLite knowledge index. Existing OpenAI and Doubao text embedding behavior is also covered.

The pre-commit checks and dashboard build passed. Provider responses were mocked; no live Ark API validation was performed.

Checklist

• Updated CHANGELOG.md (if user-facing)
• README / docs updated (if needed) — not required for this fix