fix(core): refine small model defaults

[rekram1-node]

Summary

• fall back to the active model for both Azure and Azure Cognitive Services instead of assuming a catalog small-model deployment exists
• select inferred legacy small models only from explicit models.dev families and choose the latest release in that family
• prioritize exact gemini-flash, gpt-nano, and claude-haiku families, plus gpt-mini for GitHub Copilot, with Haiku last
• avoid selecting unrelated low-capacity families such as glm-flash from Vercel AI Gateway
• preserve the existing V2 Core catalog selection algorithm, adding only the Azure provider exclusion
• document that these provider-specific assumptions should be removed once model syncing is more reliable

The current Google Vertex adapter already combines adjacent user messages into one contents entry, so this PR leaves title message construction unchanged.

Fixes #31879

Testing

• relying on CI for the latest changes at user request to avoid local CPU usage
• manually confirmed the Vertex title request combines the instruction and prompt into one user content and returns HTTP 200