Add epsilon to MTER progress postfix to prevent ZeroDivisionError (#723)

When every sampled BPR pair is skipped (e.g. on very small datasets where
all sampled pairs have equal ratings), n_bpr_samples - skipped is 0 and
computing the "correct" percentage raises ZeroDivisionError. Add + 1e-8
to the denominator, consistent with the other BPR-based models fixed in
#677 and #686.

Co-authored-by: Claude Opus 5 <noreply@anthropic.com>

Author: Shxiao

cornac/models/mter/recom_mter.pyx

@@ -419,7 +419,7 @@ class MTER(Recommender):
                 progress.set_postfix({
                     "loss": "%.2f" % (loss / 3 / self.n_element_samples),
                     "bpr_loss": "%.2f" % (bpr_loss / self.n_bpr_samples),
-                    "correct": "%.2f%%" % (100.0 * correct / (self.n_bpr_samples - skipped)),
+                    "correct": "%.2f%%" % (100.0 * correct / (self.n_bpr_samples - skipped + 1e-8)),
                     "skipped": "%.2f%%" % (100.0 * skipped / self.n_bpr_samples)
                 })