fix: route embedding requests with a supported encoding format

[z125316840-code]

Summary

• detect embedding models using LiteLLM metadata, with a model-name fallback for custom providers
• test embedding models with litellm.embedding() and explicit encoding_format="float"
• send the same explicit float encoding format for real knowledge-base dimension probes and batch embeddings
• keep chat models on litellm.completion()
• add regression coverage for connection tests and knowledge embedding calls

Root cause

FlowModel.test_connection() originally sent embedding-only models to the chat-completions endpoint. After fixing that route, real knowledge-base embedding calls still omitted encoding_format. With LiteLLM 1.83.7, the OpenAI-compatible provider received an empty value and rejected it because it only supports float or base64.

Verification

• git diff --check
• Python 3.14 bytecode compilation in the deployment image
• 25 Flow Model integration tests passed in the original fix
• 131 Flow knowledge integration tests passed with the follow-up fix
• live openai/text-embedding-v4 knowledge dimension probe returned 1024
• live deployment returned HTTP 200 with no recent severe service logs

Closes #96