Refactor object detection in valor_core to improve speed

[ntlind]

Improvements

• Add back a faster IOU calculation for axis-aligned bboxes
• Add pandas[performance] as a dependency. Use @numba.jit to speed-up IOU calculations
• Replace apply and lambda calls where possible to reduce runtime
• Move an expected output fixture from _inputs to _outputs

Performance

Base OD evaluation for AnnotationType.Box is down to ~11 seconds on the GH runner (previously ~21), and ~6 seconds locally