feat: add ark multimodal polling plugins

[WH-2099]

Summary

• Add Ark multimodal polling support to the existing VolcEngine model plugin.
• Add a separate BytePlus ModelArk plugin package instead of using multiple providers in one model plugin.
• Add Seedream 5.0 image generation and Seedance 2.0 / 1.5 Pro video generation model definitions.
• Add polling unit coverage for payload mapping, task status mapping, output conversion, provider registration isolation, and retryable check errors.

Blocked

This draft PR should not be merged until dify has completed the matching graphon and dify-plugin-daemon upgrades. The end-to-end stack needs polling invocation support before these plugin changes can be safely released.

Tracked by #3194.

Validation

• cd models/volcengine && uv run pytest
  • 24 passed, 20 skipped
• cd models/byteplus && uv run pytest
  • 17 passed, 1 skipped

Live BytePlus generation was not completed because the supplied key returned an inactive-key 401 from the provider.

[gemini-code-assist]

Code Review

This pull request introduces a new plugin for BytePlus ModelArk and updates the Volcengine Ark plugin to support multimodal models, specifically Seedance for video and Seedream for image generation. The implementation includes a polling mechanism for asynchronous tasks and logic for handling multimodal inputs like images, videos, and audio. Feedback identifies an issue in the get_num_tokens method for both plugins, where complex message content is not correctly handled during token estimation, and suggests using the _extract_prompt_text helper to fix this.