fix: test embedding models through embeddings endpoint

[z125316840-code]

Summary

• detect embedding models using LiteLLM metadata
• fall back to the model name for custom OpenAI-compatible embedding models missing from LiteLLM's registry
• test embedding models with litellm.embedding() and explicit encoding_format="float"
• keep chat models on litellm.completion()
• add regression coverage for mapped/unmapped embeddings and chat models

Root cause

FlowModel.test_connection() always sent a chat completion request. Embedding-only models therefore hit /chat/completions and returned a misleading 404 even when their model, credentials, and provider routing were valid. With the OpenAI-compatible provider used for live verification, omitting encoding_format also caused LiteLLM 1.83.7 to send an empty value rejected by the provider; using the standard float format resolves that compatibility issue.

Verification

• git diff --check
• Python 3.14 bytecode compilation in the deployment image
• 25 Flow Model integration tests passed
• live deployment verified against a custom OpenAI-compatible text-embedding-v4 model: Connection OK
• chat-model regression verifies litellm.completion() remains unchanged

Closes #96