Implement custom_ocr strategy — MiniCPM-OCR as the reference use case

[lottshin]

Problem

Uploading a scanned PDF fails with No text found in document (#17085). OCR is wired to Mistral's /ocr API (mistral_ocr, azure_mistral_ocr, vertexai_mistral_ocr) or the built-in document_parser. OCRStrategy.CUSTOM_OCR is already declared in packages/data-provider/src/config.ts, but getStrategyFunctions() in api/server/services/Files/strategies.js has no branch for it, so selecting it throws Invalid file source. That leaves self-hosted and air-gapped deployments (see also #17644) unable to use a local OCR model.

Proposal

Complete the custom_ocr strategy along the lines of #13431 (a module loaded from ocr.customStrategyModule), with a MiniCPM-OCR module as the reference implementation.

MiniCPM-OCR 1.0 is a 1.3B Apache-2.0 OCR model served through an OpenAI-compatible chat-completions endpoint, which makes it a realistic local backend for that reference module.

Why the plugin point rather than a new enum value

The existing strategies name API shapes, not models. A minicpm_ocr entry would invite one enum value per OCR model, so the pluggable path looks like the right home. #13431 has the design but is currently out of date with dev (it needs a rebase) and has had no maintainer review since June; I am happy to revive it with the author's agreement, or to contribute the reference module on top of it if that is preferred.

Open design question: PDF input

Mistral's /ocr accepts the file directly, while MiniCPM-OCR needs page images. Rendering PDF pages server-side needs a rasterizer — pdfjs-dist is already a dependency, but rendering to an image needs a canvas implementation, so this likely means a new dependency or an images-only constraint at first. Guidance on which direction maintainers prefer would be useful before writing code.

Scope I can implement

• custom_ocr handler wiring in strategies.js (or the equivalent if 🔌 feat: Pluggable Custom OCR Strategy #13431's shape changes).
• Schema and validation for the new ocr fields (customStrategyModule); ocr.baseURL and ocr.apiKey already exist and would be reused. Tests included.
• A MiniCPM-OCR module: page image → POST {ocr.baseURL}/v1/chat/completions, image content part before the prompt, text read from choices[0].message.content. Its location can follow whatever the repo prefers — there is no examples/ directory today, so a docs snippet plus a documented module may be a better fit than adding one.
• Failure semantics: backend errors and timeouts surface as upload errors; no silent fallback to document_parser unless that behavior is wanted.
• Unit tests, a librechat.example.yaml snippet, and documentation.
• I will open an issue to claim the work and wait for assignment before opening a pull request against dev, per the contributing guidelines.

Related: #13431, #17085, #17644.