Add --batch-size CLI option and document batch=1 as optimal

Fix repo name typo: rf-dert → rf-detr

Reproduce all README benchmarks with real test images

Add acknowledgments for timnielen's prior work

Clean up test scripts and fix README

Explain why ANE is unused: FP32 precision, not transformer arch

Add hardware acceleration analysis with MLComputePlan evidence

Polish for open-source release

Update README with benchmark data and English translation

fix: num_classes detection for custom weights (subtract background class)

Initial release: RF-DETR to CoreML export tool