Add multi-pass/hybrid model evaluation and image analysis when multiple models are available

[remlaps]

For example, gemma3-4b in Google's free tier. As suggested by Gemini:

Given your focus on curation—which requires high throughput to scan the blockchain but enough intelligence to detect quality, spam, and sentiment—here is the breakdown of why the 4B model is your "Goldilocks" choice, along with how the others fit in.

and

Hybrid Architecture: Use gemma-3-1b to scan the "firehose" of all incoming Steem posts to discard obvious spam (low compute cost). Pass the surviving posts to gemma-3-4b to actually score them for quality, check image relevance, and assign a vote weight.