fix: correctly list chat-capable Ollama models

[saksham-jain177]

Fixes empty model selector when Ollama returns models.

• The selector was empty due to over-filtering based on brittle metadata
  (family and parameter_size format).
• Ollama now returns values like "8.0B"
  and includes additional model types.

This change makes model discovery robust to current Ollama metadata and
excludes known non-chat models (embeddings, CLIP).

Fixes #2