Flatten q-point indices for compatibility with Numpy 2.0

With Numpy 2.0, np.unique returns a 2-D array when applied to
q-points. This in turn gets broadcast and adds an extra dimension to
the frequency/mode arrays, causing all kinds of chaos.

Flattening the array here should be a no-op for Numpy <2 and fixes the
issue.

euphonic/force_constants.py

@@ -508,6 +508,8 @@ def _calculate_phonons_at_qpts(
             norm_qpts[gamma_i] = 0.
             reduced_qpts, qpts_i = np.unique(norm_qpts, return_inverse=True,
                                              axis=0)
+            qpts_i = qpts_i.flatten()
+
             n_rqpts = len(reduced_qpts)
             # Special handling of gamma points - don't reduce gamma
             # points if LO-TO splitting