feat: Add DOCLING_SERVE_DEFAULT_OCR_ENGINE environment variable for deploy-level OCR engine configuration

## Summary

There is currently no way to configure the default OCR engine at deploy time. The `ocr_engine` parameter only exists at the request level, meaning every client must explicitly pass `ocr_engine=tesseract` (or whichever engine) on every request, or they silently get EasyOCR regardless of the server environment.

## Expected behaviour

A new environment variable — `DOCLING_SERVE_DEFAULT_OCR_ENGINE` — should allow operators to set the OCR engine server-wide at deploy time, so that requests which do not specify an `ocr_engine` fall back to the configured default rather than always defaulting to EasyOCR.

```bash
DOCLING_SERVE_DEFAULT_OCR_ENGINE=tesseract
```

## Motivation

- In many deployment environments (CPU-only, FIPS-compliant, lightweight containers), EasyOCR is undesirable or broken, while Tesseract is the preferred engine.
- The pattern already exists in the codebase for other pipeline components — `DOCLING_SERVE_DEFAULT_TABLE_STRUCTURE_KIND` and default layout kind are both configurable at the server level. OCR is a notable omission.
- Forcing every downstream client to pass `ocr_engine` on every request is not practical when docling-serve is used as infrastructure (e.g. behind Open WebUI or other integrations that don't expose this parameter).

## Suggested implementation

Follow the same pattern as the existing default table/layout settings in `docling_serve/settings.py` — add a `default_ocr_engine` field to `DoclingServeSettings` with `DOCLING_SERVE_DEFAULT_OCR_ENGINE` as its env var, and apply it as the fallback when no `ocr_engine` is provided in the request options.

Start in docling_serve/settings.py and compare the existing default table-structure and layout settings. Trace how request options apply their OCR engine fallback. Done means DOCLING_SERVE_DEFAULT_OCR_ENGINE configures the server-wide default while an explicitly provided request value still takes precedence.

Written by the indexing model from the issue text.