fix: price OpenAI requests by actual service tier

[dgageot]

Why

OpenAI's fast/priority/ultrafast service tiers charge a premium over
the base model rate, but that premium wasn't tracked per-model: it was a
hardcoded list of model names that would silently stop applying once a model
rolled off the allowlist, and never covered models with a different premium
than 2x/6x.

What

• Replaced the model-name allowlist in Cost.ForServiceTier with a
  fastMultiplier derived per-model from the catalogue (withFastPricing in
  pkg/modelsdev/service_tier.go), looked up once in Store.GetModel.
• The multiplier is only trusted when the catalogue's base rates match our
  own and the premium is uniform across input/output/cache rates and all
  context tiers; otherwise it falls back to the previous flat estimates
  (fast = 2x, ultrafast = 6x), as before.
• Dated model snapshots (e.g. -2025-01-01 suffix) still fall back to the
  alias's premium, but no longer assume the alias's base rates apply to them.
• Endpoint/override/cache/context pricing behavior is unchanged.
• Updated docs and the examples/openai-service-tier.yaml example, and
  extended tests in pkg/modelsdev and pkg/runtime.

Validation

• CI=true task build, task test, task lint
• Targeted -race runs for the modelsdev/runtime service-tier tests and
  the catalogue boundary tests
• Manual smoke test against live gpt-6.1-sol ultrafast pricing