[FEATURE]: Local OCR fallback for images when no vision model is available

Feature hasn't been suggested before.

• I have verified this feature I'm about to request hasn't been suggested before.

Describe the enhancement you want to request

Follow-up to the image transcription feature (#48332). When no vision-capable model exists in the selected model's provider, images are still not readable.

This is the same goal as #46785 (RapidOCR-json) but with a bundled Tesseract engine instead of a separately installed OCR binary, so it works out of the box.

Request: a fully local OCR fallback that runs on the user's machine, so the image never leaves the device.

Findings from a prototype:

• tesseract.js does not work in the compiled bun build --compile binary. The worker thread cannot resolve wasm-feature-detect and other modules because the on-disk node_modules paths are not present in the binary. The build succeeds but createWorker hangs.
• tesseract-wasm is a better fit: its core WASM is ~1.8MB (vs ~43MB for tesseract.js-core), it exposes a synchronous createOCREngine({ wasmBinary }), and it worked in a source run. Its WASM is not exported from the package exports, so embedding it in the binary needs either vendoring the WASM or a build step that copies it next to the binary.
• Language data (eng.traineddata.gz, ~14MB) can be bundled and loaded with loadModel, no network needed.
• Image input needs decoding to RGBA before loadImage; photon-node is already a dependency and can decode.

Open questions: where to put the OCR engine (main thread vs worker), how to embed the WASM in the compiled binary, and whether OCR should be opt-in via config.