fix(ocr): clamp PDF image bboxes before cropping

[cyphercodes]

Summary

• Clamp PDF image crop bboxes to the pdfplumber page bbox before calling within_bbox.
• Keep OCR extraction working when image metadata has tiny floating-point overflows, such as a slightly negative top value.
• Add a focused regression test for the out-of-bounds bbox case.

Fixes #2097

Tests

• pytest -q packages/markitdown-ocr/tests/test_pdf_converter.py::test_extract_images_clamps_slightly_out_of_bounds_bbox
• pytest -q packages/markitdown-ocr/tests/test_pdf_converter.py -k 'not test_pdf_multipage'
• git diff --check

Note: running the full packages/markitdown-ocr/tests/test_pdf_converter.py currently fails on test_pdf_multipage with the current dependency set; I reproduced the same failure on unmodified main before re-applying this patch.