Add ONNX CPU benchmarking

[leeclemnet]

What does this PR do?

Adds SAB support for ONNX CPU execution provider. This is useful for edge device profiling and particularly for benchmarking CPU-optimized models like YOLOLite-Edge variants.

• Extracted ProfilerBase in profiler.py, added CPUProfiler — uses time.perf_counter_ns() for wall-clock timing on CPU
• Refactored ONNXInference into ONNXInferenceBase → ONNXInferenceCUDA / ONNXInferenceCPU in onnx_inference.py
• Updated pretty_print_results() — runtime column shows ONNX-CPU / ONNX-CUDA / TRT
• Added is_cpu to dump() output for result serialization
• Added RFDETRONNXCPUInference in benchmark_rfdetr.py with CPU benchmark requests for nano, small, and medium variants (grouped after TRT requests per model size)
• Add ONNXInference warmup step and fix intra/inter_op thread count for stable CPU inference timing

Type of Change

• New feature (non-breaking change that adds functionality)

Testing

• I have tested this change locally
• I have added/updated tests for this change

Test details:

Tested benchmark_rfdetr.py against RF100 aquarium test set on n1-standard-16 VM + T4.

python -m sab.models.benchmark_rfdetr rf100/aquarium-qlnqy/test rf100/aquarium-qlnqy/test/_annotations.coco.json

Model                          Runtime  FP16   mAP50  mAP50-95   AP75   Latency Throttled
-----------------------------------------------------------------------------------------
rf-detr-nano.onnx              TRT      no       4.7       3.2    3.4      9.84       yes
rf-detr-nano.onnx              TRT      yes      4.9       3.3    3.6      2.36        no
rf-detr-nano.onnx              ONNX-CPU no       4.7       3.2    3.4     99.99        no
rf-detr-small.onnx             TRT      no       3.4       2.5    2.9     18.99       yes
rf-detr-small.onnx             TRT      yes      3.5       2.6    2.9      3.60       yes
rf-detr-small.onnx             ONNX-CPU no       3.4       2.5    2.9    177.84        no
rf-detr-medium.onnx            TRT      no       6.2       4.8    5.4     25.88       yes
rf-detr-medium.onnx            TRT      yes      6.1       4.7    5.5      4.65       yes
rf-detr-medium.onnx            ONNX-CPU no       6.2       4.7    5.3    233.72        no

AP breakdown (COCO):
Model                            AP_s   AP_m   AP_l
---------------------------------------------------
rf-detr-nano.onnx                 0.1    1.3    4.6
rf-detr-nano.onnx                 0.2    1.3    4.8
rf-detr-nano.onnx                 0.1    1.3    4.6
rf-detr-small.onnx                0.1    1.1    3.8
rf-detr-small.onnx                0.1    1.2    3.8
rf-detr-small.onnx                0.1    1.1    3.8
rf-detr-medium.onnx               0.3    4.4    6.7
rf-detr-medium.onnx               0.3    4.4    6.5
rf-detr-medium.onnx               0.3    4.3    6.7

AR breakdown (COCO):
Model                            AR@1  AR@10   AR@max_dets   AR_s   AR_m   AR_l
-------------------------------------------------------------------------------
rf-detr-nano.onnx                 4.2   11.1          13.7    4.2   10.1   20.1
rf-detr-nano.onnx                 4.2   12.1          14.7    4.6   10.1   22.1
rf-detr-nano.onnx                 4.2   11.1          13.7    4.2   10.1   20.1
rf-detr-small.onnx                5.0   12.0          14.6    3.9    8.4   23.1
rf-detr-small.onnx                5.0   12.1          14.5    3.9    8.3   22.4
rf-detr-small.onnx                5.0   12.0          14.6    3.9    8.4   23.1
rf-detr-medium.onnx               6.4   18.0          20.6    5.4   15.9   28.8
rf-detr-medium.onnx               6.0   18.0          20.6    5.3   16.2   26.7
rf-detr-medium.onnx               6.4   17.7          20.2    5.4   15.1   28.8

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code where necessary, particularly in hard-to-understand areas
• My changes generate no new warnings or errors
• I have updated the documentation accordingly (if applicable)

Additional Context