[cerebras] Refresh public model catalog

[rekram1-node]

Summary

• reconcile Cerebras with its unauthenticated first-party public model catalog (2 active models)
• add provider-verified reasoning controls: GPT OSS low|medium|high effort and GLM 4.7 none effort
• correct current pricing, output limits, structured-output support, GLM reasoning/preview metadata, and remove unavailable llama3.1-8b

Control matrix

Model	Reasoning	Verified normalized control	Provider contract
gpt-oss-120b	yes	effort: low, medium, high	reasoning_effort; medium default
zai-glm-4.7	yes	effort: none	reasoning enabled by default; reasoning_effort=none disables

The deprecated GLM disable_reasoning boolean is not used as evidence for a normalized toggle. The active documented wire control is reasoning_effort, for which Cerebras documents only the none value on GLM 4.7.

Official evidence

• https://api.cerebras.ai/public/v1/models
• https://inference-docs.cerebras.ai/api-reference/chat-completions
• https://inference-docs.cerebras.ai/capabilities/reasoning
• https://inference-docs.cerebras.ai/models/overview
• https://inference-docs.cerebras.ai/support/deprecation

Local validation

• bun validate
• git diff --check
• no test files are included in this PR
• no CEREBRAS_API_KEY was available, so no authenticated inference probes were run; no controls rely on upstream behavior or generic OpenAI compatibility