Normalize MMR diversity term for consistent behavior across distance metrics

The MMR relevance term is normalized to [0,1] by dividing by max_dist,
but the diversity term used raw distances (1 - d). For cosine distance
this works fine since values are already in [0,1], but for L2 and L1
where distances are unbounded, the two terms in the MMR score operated
on different scales, making mmr_lambda behave unpredictably.

Normalize the diversity term the same way (d / max_dist) so both terms
are on a consistent [0,1] scale regardless of the distance metric.

Author: vlasky

sqlite-vec.c

@@ -7624,7 +7624,7 @@ static int vec0_mmr_rerank(
             for (i64 j = 0; j < step; j++) {
                 f32 d = vec0_compute_distance(vector_column,
                                               vectors[i], out_vectors[j]);
-                f32 sim = 1.0f - d;
+                f32 sim = 1.0f - (d / max_dist);
                 if (sim > max_sim) max_sim = sim;
             }
 

tests/__snapshots__/test-knn-mmr.ambr

@@ -192,14 +192,14 @@
         'rowid': 1,
         'distance': 0.0,
       }),
-      OrderedDict({
-        'rowid': 3,
-        'distance': 0.028284257277846336,
-      }),
       OrderedDict({
         'rowid': 2,
         'distance': 0.014142128638923168,
       }),
+      OrderedDict({
+        'rowid': 5,
+        'distance': 1.4142135381698608,
+      }),
     ]),
   })
 # ---