Surface gateway controls (skip-cache, request metadata) and per-model cost; add Dynamic Routing

Problem/Motivation

The Cloudflare AI Gateway provider only wired the cache_ttl of the cf-aig-* controls the gateway request value object already models. It surfaced no per-model cost, did not make Cloudflare Dynamic Routing discoverable, had no explicit test guarding the enabled-but-unconfigured state, and streamed chat fataled for Workers AI models.

Proposed resolution

• Wire cf-aig-skip-cache and cf-aig-metadata into requests. Metadata combines an operator-configured key/value map with the AI operation type and any per-request tags, so requests are filterable in the Cloudflare dashboard. Adds matching settings-form fields and config schema. The OpenAI client is rebuilt per request so per-request metadata reflects the current call rather than pinning the first call's tags for the instance's life.
• Add a ModelLabel helper that appends each model's owner and per-million-token cost to the autocomplete and configured-model lists, de-duplicating the label logic shared by the controller and the provider. The /compat/models catalogue prices per token for most models but per million for some legacy Workers AI entries, so the label normalizes both to a per-million figure.
• Document dynamic/<route> models (Cloudflare Dynamic Routing) in the chat models help text, with a characterization test that they reach the compat endpoint unmangled.
• Fix streamed chat for Workers AI models, whose compat streaming serializes a numeric token as a JSON number ({"delta":{"content":1}}) and trips openai-php's strict ?string content: a response-body stream decorator rewrites each SSE data: line's numeric content to a string in real time, without buffering the stream.
• Add a regression test that the provider degrades without throwing when no gateway is configured, guarding the enabled-but-unconfigured settings-page fatal (core ai #3572645 class).

Remaining tasks

Review and merge.