Install onnxruntime-gpu for non-Darwin platform

[wkentaro]

So that it can accelerate with CUDA when possible.

cc @knorth55

[knorth55]

I confirmed that this is working as I expected.

CPU

$ python spam.py
2024-02-12 10:02:09.302 | INFO     | osam._models._base:__init__:91 - Initialized inference sessions with providers ['CPUExecutionProvider']
2024-02-12 10:02:09.302 | WARNING  | osam.apis:generate:31 - Prompt is not given, so using the center point as prompt: Prompt(points=array([[960., 720.]], dtype=float32), point_labels=array([1], dtype=int32))
Duration: 3.131537675857544
2024-02-12 10:02:11.647 | WARNING  | osam.apis:generate:31 - Prompt is not given, so using the center point as prompt: Prompt(points=array([[960., 720.]], dtype=float32), point_labels=array([1], dtype=int32))
Duration: 2.272653341293335

CUDA

$ python spam.py
2024-02-12 10:02:26.820 | INFO     | osam._models._base:__init__:91 - Initialized inference sessions with providers ['CUDAExecutionProvider', 'CPUExecutionProvider']
2024-02-12 10:02:26.821 | WARNING  | osam.apis:generate:31 - Prompt is not given, so using the center point as prompt: Prompt(points=array([[960., 720.]], dtype=float32), point_labels=array([1], dtype=int32))
Duration: 2.08842396736145
2024-02-12 10:02:28.011 | WARNING  | osam.apis:generate:31 - Prompt is not given, so using the center point as prompt: Prompt(points=array([[960., 720.]], dtype=float32), point_labels=array([1], dtype=int32))
Duration: 0.40121889114379883

[wkentaro]

Thanks for checking @knorth55

FYI, I added benchmark.py in examples folder:
https://github.com/wkentaro/osam/blob/main/examples/benchmark/benchmark.py