docs: correct claims that measurement did not support

- model-router: README said ~180ms per routing decision. Measured 210-460ms
  round trip over ~40 calls from a laptop; no call came in under 200ms. Replaced
  the number with the measured range and said plainly that most of it is network.
- model-router/multimodel: the provider env var names are this example's own
  convention, not something the Morph API reads, and Morph does not proxy the
  router's catalog (glm-5.2, claude-opus-4-8 etc. 404 on api.morphllm.com).
  Both now stated instead of implied.
- open-source-models: dropped the morph-qwen35-397b row. Requests for it are
  served by morph-qwen36-27b (reproduced 5/5, response.model comes back
  morph-qwen36-27b every time), so advertising a 397B model with 262k context
  would be wrong. Restore the row once routing is fixed upstream.