fix: support omitDimensions for local embedding models

[yjjheizhu]

Summary

This PR is a clean replacement for #264 and keeps only the omitDimensions compatibility fix for local embedding models such as Qwen3/vLLM/Ollama.

What changed

• add embedding.omitDimensions to the plugin config schema and UI metadata
• parse and pass embedding.omitDimensions through plugin config loading
• omit the dimensions field from embedding requests when omitDimensions: true
• add a regression test to ensure the plugin manifest/schema stays in sync

Why

Some OpenAI-compatible local embedding endpoints reject requests that include the dimensions parameter. This makes the existing omitDimensions configuration necessary for compatibility, but it must be wired through config parsing and request construction to actually work.

Testing

• node test/embedder-error-hints.test.mjs
• node test/plugin-manifest-regression.mjs

Notes

• This intentionally does not include the unrelated extra changes currently mixed into fix: local embedding model compatibility (Qwen3/vLLM/Ollama) + configurable rerank timeout #264.
• upstream/master currently has an unrelated failing regression in test/cjk-recursion-regression.test.mjs; this PR does not modify that area.