[Feature] Allow model size configuration for RapidOCR

### Requested feature
PP-OCRv6 has `tiny` and `medium` model sizes, which are [supported by RapidOCR](https://www.modelscope.cn/models/RapidAI/RapidOCR/tree/master/onnx/PP-OCRv6/det), but Docling currently [hardcodes model size `small`](https://github.com/docling-project/docling/blob/v2.120.2/docling/models/stages/ocr/rapid_ocr_model.py#L34). An option to change that setting would be great.

From my profilings on a CPU-only scenario, OCR is the slowest stage by far, even on PDF files with native text mostly avaliable. Maybe the `tiny` model could help improve this.

Start in docling/models/stages/ocr/rapid_ocr_model.py, where the PP-OCRv6 model size is hardcoded to small, and trace how RapidOCR model settings are configured. Add a user-selectable size supporting the tiny and medium models documented in the issue, then verify OCR can use each supported size.

Written by the indexing model from the issue text.