Simplify conversion process by only converting weights

[decahedron1]

The computation graph for all Stable Diffusion v1, v2, and v2.1 models are constant. The only thing that changes are the actual weights, but the current conversion process requires doing a full graph capture via torch.onnx.export, which takes a long time and uses a lot of memory. Additionally, models are not as well optimized as they could be (partly because running onnxsim on the UNet uses an impractical amount of memory).

Instead of converting the graph every time, we could instead provide a "skeleton" ONNX graph (potentially using the external data format?) which the conversion script adds the model weights to. This makes model conversion much faster and allows us to provide manual graph optimizations.