Currently sznn only supports a small subset of ONNX operators.
All pull requests of operators implementations are welcome.

More threads not always bring more performance benifits.
We should determine thread number by the scale of the data.

Because sznn only implements a part of the operators, we cannot offer objection detection example.
Hope in the future when more operators are added, we can present YOLO as detection example.

For convenience, current implementation of WebGPU backend copys data every time when executing a operator.
HPC is all about data movement. It's hard to imagine how much time we spent on these useless data movement.
Consider replace data field with WebGPU buffer type member in WebGPU tensor definition.
And computePass() function should be modified as well, never copy back result back in its end.

Convolutional operator is the most time-consuming operator in CNN due to statistics.
so it is worth optimizing its implementation.
For small size kernel we can use imcol + winograd algorithm, while for the large size kernel we can use FFT trick and so on.