fix(deepinfra): reconcile reasoning controls

[rekram1-node]

Summary

• audit every configured DeepInfra model against the live catalog, official reasoning documentation, OpenAPI schema, and model API pages
• expose DeepInfra's documented reasoning toggle and low/medium/high effort controls on all reasoning models
• remove reasoning options from non-reasoning models and remove undocumented xhigh values

Validation

• bun validate
• git diff --check
• resolved DeepInfra invariant audit: 25 models, 21 reasoning, 4 non-reasoning, no invalid option combinations
• bun test packages/core/test/generate.test.ts: 9 pass, 1 unrelated repository-wide failure for missing weights links in NVIDIA/Sarvam/Cohere metadata

Audit notes

• all 25 configured models remain present in DeepInfra's public model catalog
• no DeepInfra API credential was available, so authenticated behavior testing was not performed
• no token-budget control is documented by DeepInfra
• prior DeepInfra PR feat(deepinfra): add reasoning options #2083 is merged; no open DeepInfra PR existed before this PR