Fix Workers AI model reasoning input types - #7437

Refresh Workers AI binding types to describe each model's reasoning efforts and whether thinking can be disabled. Previously, shared low | medium | high types rejected valid controls such as Gemma's none and GLM's max. The updated types include canonical efforts and accepted normalization aliases, while mandatory-reasoning models reject disabling thinking. Shared type defaults remain compatible for models without metadata.

Regenerated the reasoning declarations, added the missing GLM-5.3 binding entry, and preserved unrelated input fields and token limits. The accompanying documentation update explains canonical controls separately from aliases.

Validation: generated type snapshots and focused compile-time coverage for Chat Completions, Responses, aliases, invalid efforts, mandatory thinking, and unchanged options.