fix(think): restore workspace image reads on AI SDK 6 - #2082

This PR restores image-specific model output for Think workspace reads across supported AI SDK peers. Fixes #2080.

Why

• @cloudflare/think supports ai@6 and ai@7, but workspace images currently emit file-data, the same content part used for PDFs.
• AI SDK 6 treats image-data as image input and file-data as document input. Images therefore fail or get dropped before reaching the model.
• AI SDK 7 normalizes either legacy part using its media type, which hid the regression.
• We could detect the installed AI SDK major and emit its preferred shape, but runtime version detection adds complexity without changing behavior. image-data preserves image semantics on both supported majors.

Code Changes

• Restore the image-specific image-data branch in the workspace read tool’s model-output conversion.
• Keep PDFs on the existing file-data path with their filename.
• Preserve compact workspace results and rehydrate their bytes only when preparing model output.
• Add an @cloudflare/think patch changeset.

Compatibility

• AI SDK 6 now routes workspace image reads as images rather than documents.
• AI SDK 7 accepts and normalizes image-data, preserving its existing provider-facing image behavior.
• Consumers do not need to update call sites.