Fix NumPy-safe coefficient checks

[rogeriojorge]

Summary

• replace empty-list equality checks for coefficient sequences with length-based checks that also work for NumPy arrays
• fix the three accidental comparison expressions that prevented d_over_curvature_cvals from being assigned
• initialize or pad the automatic second-order coefficients before building the optimization parameter list, without adding duplicate parameters
• add regression coverage for set_dofs(), scalar initialization, automatic construction, padding, and NumPy coefficient inputs

Root cause

get_dofs() returns a NumPy vector, and set_dofs() assigns slices of that vector back to coefficient attributes. A subsequent calculation then evaluated expressions such as self.d_over_curvature_cvals == [], which attempt incompatible elementwise broadcasting for a non-empty NumPy array.

The assignment typos in Qic.__init__() and construct_qi_r2() also meant that generated coefficient arrays were silently discarded.

Testing

• python -m unittest -v qic.tests.test_d_over_curvature on Python 3.11: 5 passed
• python -m unittest -v qic.tests.test_d_over_curvature on Python 3.12 with NumPy 2.5.2: 5 passed
• exact issue reproducer: passes and preserves both coefficients and d
• documented first construct_qi_r2() path: completes as order r2 with B2cQI_deviation_max = 3.2580160791439994e-10
• full headless suite: 35 tests run with the same three pre-existing numerical-reference failures and one missing-mpi4py import error seen on pristine main; no new failures