Operation decomposition should be done in ModelBuilder

Some operations in DNNLibrary are indeed composited by several nnapi operations

e.g.,
dilated convolution = space to batch -> convolution -> batch to space
prelu -> relu + (mul -> relu -> mul).

The decomposition is now done in OnnxConverter, however, many of these operations are supported directly by nnapi in Android Q. As a result, the decomposition should be done in ModelBuilder on-the-fly in order to choose the best interpretation of these ops.