feat: benchmark app improvements

[mgumowsk]

What does this PR do?

Fixes #418

Change overview

• Adds options to specify threads, streams, performance_hint to benchmark.app
• Reporting in ms
• Async support

Before changes

uv run python examples/metrics/benchmark.py   data/otx_2.6.0_models/detection_rtmdet_tiny_fp16/openvino.xml   data/coco128/images/train2017   --device CPU  --test-runs 100
Found 128 images in data/coco128/images/train2017
OpenVINO Runtime
	build: 2025.3.0-19807-44526285f24-releases/2025/3
Reading model data/otx_2.6.0_models/detection_rtmdet_tiny_fp16/openvino.xml
The model data/otx_2.6.0_models/detection_rtmdet_tiny_fp16/openvino.xml is loaded to CPU
	Number of model infer requests: 2
The model data/otx_2.6.0_models/detection_rtmdet_tiny_fp16/openvino.xml is loaded to CPU
	Number of model infer requests: 2
Starting warm-up...
Running 100 test inferences...
  Completed 10/100
  Completed 20/100
  Completed 30/100
  Completed 40/100
  Completed 50/100
  Completed 60/100
  Completed 70/100
  Completed 80/100
  Completed 90/100
  Completed 100/100

============================================================
               🚀 PERFORMANCE METRICS REPORT 🚀
============================================================

📊 Model Loading:
   Load Time: 0.282s

⚙️  Processing Times (mean ± std):
   Preprocess:  0.000s ± 0.000s
   Inference:   0.023s ± 0.002s
   Postprocess: 0.000s ± 0.000s

📈 Total Time Statistics:
   Mean:  0.023s ± 0.002s
   Min:   0.021s
   Max:   0.035s

🎯 Performance Summary:
   Total Frames: 100
   FPS:          43.36

============================================================

After changes

uv run python examples/metrics/benchmark.py   data/otx_2.6.0_models/detection_rtmdet_tiny_fp16/openvino.xml   data/coco128/images/train2017   --device CPU --performance-hint THROUGHPUT --test-runs 100 --async --inference-only --max-num-requests `nproc` --num-streams `nproc`
Reading model data/otx_2.6.0_models/detection_rtmdet_tiny_fp16/openvino.xml
The model data/otx_2.6.0_models/detection_rtmdet_tiny_fp16/openvino.xml is loaded to CPU
	Number of model infer requests: 16
The model data/otx_2.6.0_models/detection_rtmdet_tiny_fp16/openvino.xml is loaded to CPU
	Number of model infer requests: 16
Benchmarking in inference only mode (inputs filling are not included in measurement loop).
Input: image, shape: [1, 3, 640, 640], precision: f32, layout: NCHW
Compiled model input partial_shape: [1,?,?,3], layout: NCHW
Using tensor shape: [1, 224, 224, 3]
Raw tensor shape: (1, 224, 224, 3), dtype: float32
Starting warm-up...
Running 100 test inferences (async mode, inference only)...
  Submitted 10/100
  Submitted 20/100
  Submitted 30/100
  Submitted 40/100
  Submitted 50/100
  Submitted 60/100
  Submitted 70/100
  Submitted 80/100
  Submitted 90/100
  Submitted 100/100
  Waiting for all requests to complete...
  Completed 100/100

============================================================
           🚀 PERFORMANCE METRICS REPORT 🚀
============================================================

📊 Execution Devices: [CPU]
   Count: 100 iterations
   Duration: 1424.65 ms

📈 Latency:
   Average: 14.25 ms

🎯 Throughput: 70.19 FPS
============================================================