Derives USD cost from token usage and a model id.

Per-million-token input/output rates are stored per model. Cache rates are properties of the caching mechanism, not the model, so they are applied as multipliers on the input rate: cache reads cost 0.1x input, and cache writes cost 1.25x input (the 5-minute TTL rate). The rarer 1-hour TTL write rate (2x) is not modeled here; reads dominate cost, so the resulting estimate error is negligible. Unknown models return nil.