• Added optimization fusing an outer product and a sum over the parameters (e.g., embedding weights) into an einsum. This is crucial to avoid OOM errors and reproduce the benchmark results in the SOS circuits paper.

Other changes:

• Torch parameter pointers now hold the pointed torch tensor parameter as a submodule.
• Removed TorchCircuit.device() property and the ._set_device() method for two reasons. First, modules are NOT stored on devices, but tensors are. Second, when the circuit is a sub-module in some other module, then its device was NOT updated when calling .to() to the parent module. This is because .to() is applied recursively to the tensors (buffers & parameters) but not modules. This broke the implementation of SOS circuits.
• We now do not need to pass a dummy tensor to the circuit in order to evaluate the partition functions. That is, constant layers now only take the batch size as input, rather than some allocated dummy tensor.
• Fixed a few pylint errors.