AI Gateway [2.0]: Document dynamic multimodal LLM pricing for AI Proxy Advanced

We're introducing aew `config.pricing` record on `ai-proxy-advanced` with inline JSON catalog (`pricing.enabled`, `pricing.models`). Adds per-modality cost fields (text, audio, image, video), `cache_read_per_token`, and `cache_write_per_token`. Inline pricing overrides static `model.options.input_cost` / `output_cost`.

DoD:

* Reference update: new `config.pricing` fields, catalog entry shape, validation rules.
* Full pricing field set, including modality-specific and cache fields.
* Resolution order: inline match, then `model.options`, then zero. Pricing failures don't block requests.
* Two config examples: inline catalog enabled, and fallback path with `enabled: false`.
* Limitations: `cache_write_per_token` schema-only (not calculated), no per-modality or cached-token metrics surfaced, embeddings folded into text input, realtime reuses multimodal schema.

TBD: How this will work with AI GW 2.0

Locate the AI Proxy Advanced reference documentation and its existing configuration examples, then read how model pricing is currently described. Document the config.pricing fields, catalog shape, validation, resolution order, examples, and stated limitations; done means each DoD item is covered without resolving the AI Gateway 2.0 TBD.

Written by the indexing model from the issue text.