Support JIT for class and class method + nnf config + 3-lvl-signature

[siahuat0727]

As discussed,

Added

• Support JIT for class method (as decorator and as function)

  • class Net(...):
    @nnfusion.jit
    def forward(...):
        ...

  • 
    class Net(...):
    def forward(...):
        ...

  model = Net() model.forward = nnfusion.jit(model.forward)

  
  + Note that these two approaches are essentially different because all instances share the same JIT wrapper in the former.

• Support JIT for class
  • is equivalent to JIT forward

  • @nnfusion.jit
    class Net(...):
    def foo(...):
        ...
    def forward(...):
        ...

    will have the same behavior as

    class Net(...):
    def foo(...):
        ...
    @nnfusion.jit
    def forward(...):
        ...

• Implement 3-level signature

  • |-- object1-signature
    |   |-- model1-hash
    |   |   |-- flag1-hash
    |   |   |   `-- codegen
    |   |   |-- flag2-hash
    |   |   |   `-- codegen
    |   |   `-- model.onnx
    |   `-- model2-hash
    |       |-- flag1-hash
    |       |   `-- codegen
    |       `-- model.onnx
    `-- object2-signature

• Add nnfusion.Config to fully leverage nnfusion-cli
• Add NNFusion JIT docs

Changed

• Treat model params as inputs so that different model instances can share the same kernel.
• Return the object itself if the target object is an instance of torch.nn.Module (instead of returning a function).
• Change the first part of the object's signature from relative path to __module__

I'd love to hear your feedback.

392