openrouter: stick to observed provider for the session to reuse KV cache

Once OpenRouter actually serves a request for a model from a concrete
upstream provider, pin that provider for the rest of the session and
disable fallbacks so every subsequent turn routes to the same backend
and keeps prompt-prefix (KV) caching warm. Previously the observed pin
only held while a cache read was recent (1h TTL) and still allowed
fallbacks, so OpenRouter could re-select a cold provider mid-session.

Explicit user-configured routing order still wins; an explicit @provider
pin keeps its own fallback preference.