fix mismatch in opinion handling

[MacchiatoCreamCheese]

Description

This change fixes two mismatch in the Companion model’s user–item–aspect–negative opinion (UIAON) in recom_companion.pyx.

1. Score lookup for the negative-opinion pair
   When the sampled negative opinion o_jdx is present in user_item_aspect_neg_opinion_dict, the code compares i_score to j_score from the same (u,i,a,o) key. That j_score read from user_item_aspect_pos_opinion_dict, which is inconsistent. It now reads from user_item_aspect_neg_opinion_dict, so it are based on the correct dict now.

2. Gradients
   The UIAON prediction uses G3 (and del_g3 was already updated with G3), but the user-factor loop was backpropagating del_u / del_a / del_o with G2. Those terms now use G3, so gradients align with the tensors used.

Related Issues

N/A

Checklist:

• I have added tests.
• I have updated the documentation accordingly.
• I have updated README.md (if you are adding a new model).
• I have updated examples/README.md (if you are adding a new example).
• I have updated datasets/README.md (if you are adding a new dataset).

[lthoang]

LGTM