Support RF-DETR Triton postprocess through 2XLarge

[aseembits93]

Summary

• add tiled positive-source bounds for RF-DETR Triton sparse RLE postprocess so non-nano mask sizes can use the optimized path
• raise RF-DETR Seg sparse postprocess shape limits through 2XLarge and add focused tests
• document the Jetson Orin 1080p benchmark results showing large is the largest tested variant above 30 FPS

Tests

• PYTHONPATH=/app/helloworld/inference/inference_models:/app/helloworld/inference python -m pytest tests/unit_tests/models/rfdetr/test_triton_postprocess.py

Benchmark Notes

• rfdetr-seg-large/1 depth 2: 35.49 FPS
• rfdetr-seg-xlarge/1 depth 2: 20.94 FPS
• rfdetr-seg-xlarge/1 depth 3: 20.91 FPS
• rfdetr-seg-2xlarge/1 depth 2: 12.90 FPS