Orbital optimization uses redundant parameters

[kevinsung]

The orbital optimization routine accepts a list of parameters k_flat of length n_orbitals**2. It is then used to construct an antisymmetric matrix:

qiskit-addon-sqd/qiskit_addon_sqd/fermion.py

Lines 523 to 524 in 9b57485

	p = jnp.reshape(k_flat, (num_orbitals, num_orbitals))
	K = (p - jnp.transpose(p)) / 2.0

A real-valued antisymmetric matrix is parameterized by its upper triangle excluding the diagonal, so only n_orbitals * (n_orbitals - 1) / 2 parameters are needed. Eliminating the redundant parameters should improve the convergence of the optimization.