Support per-unit pricing (per-image, per-second) in cost

Is your feature request related to a problem?

cost is denominated per million tokens, so models that charge per image or per second have nowhere to put their price and end up with cost: {}.

xAI's published API prices, none of which the schema can hold:

OpenAI's gpt-image-1, gpt-image-1.5, gpt-image-1-mini and chatgpt-image-latest are in the same position.

Describe the solution you'd like

A per-unit price alongside the existing token fields, so text models are unaffected:

[cost.per_unit]
unit = "image"   # or "second"
amount = 0.04

Describe alternatives you've considered

#66 proposes per-modality costs with a quality × dimensions matrix. That fits OpenAI's tiering, but not xAI's flat per-image price, and it doesn't cover per-second video — so video stays unrepresentable even if it ships as written. Happy for this to be folded into #66 if you'd rather keep one thread; the per-second case is the part I didn't want to lose.

Consumers looking prices up themselves. These providers publish pricing as marketing HTML, not an API, so it would mean scraping.

Additional context

We hit this in jazz, trying to cap what an agent may spend on image generation. We ended up capping the number of calls instead — a dollar cap silently becomes no cap on exactly the models that report no price.

Start by locating the existing cost schema and token-pricing fields in api.json, then compare how current model pricing entries are represented. Done means the schema can represent per-image and per-second prices through a per_unit value while leaving token-based text model costs unchanged.