Add a custom optimizer that implements AdamW with proper magnitude computation for complex tensors

[dhpitt]

torch.optim optimizers view all complex parameters as real before computing update steps. Even after the 2021 update to support complex tensors, the magnitudes computed for momentum are thrown off by computing grad * grad.conj() for a 2-tensor real stack, which is equivalent to squaring the real and imaginary components separately. This is different than an actual multiplication of grad * grad.conj() for complex numbers:

>>> x = torch.randn((2,2), dtype=torch.cfloat)

>>> x

tensor([[-1.1954+0.7736j, -0.3455+0.1481j],
        [-1.2845+0.6057j,  0.1566+0.3581j]])

>>> y = torch.view_as_real(x)

>>> y

tensor([[[-1.1954,  0.7736],
         [-0.3455,  0.1481]],

        [[-1.2845,  0.6057],
         [ 0.1566,  0.3581]]])

>>> torch.isclose(y**2, y * y.conj()).all()

tensor(True)

>>> torch.isclose(x**2, x * x.conj()).all()

tensor(False)