OPTIM: ebtree: make ebmb_insert_prefix() keep a copy the new node's key

Similarly to the previous patch, it's better to keep a local copy of
the new node's key instead of accessing it every time. This slightly
reduces the code's size in the descent and further improves the load
time to 7.45s.
diff --git a/include/import/ebmbtree.h b/include/import/ebmbtree.h
index fc6a500..365042e 100644
--- a/include/import/ebmbtree.h
+++ b/include/import/ebmbtree.h
@@ -620,6 +620,7 @@
 	int old_node_bit;
 	unsigned int npfx = new->node.pfx;
 	unsigned int npfx1 = npfx << 1;
+	const unsigned char *nkey = new->key;
 
 	side = EB_LEFT;
 	troot = root->b[EB_LEFT];
@@ -679,13 +680,13 @@
 			/* No need to compare everything if the leaves are shorter than the new one. */
 			if (len > old->node.pfx)
 				len = old->node.pfx;
-			bit = equal_bits(new->key, old->key, bit, len);
+			bit = equal_bits(nkey, old->key, bit, len);
 			break;
 		}
 
 		/* WARNING: for the two blocks below, <bit> is counted in half-bits */
 
-		bit = equal_bits(new->key, old->key, bit, old_node_bit >> 1);
+		bit = equal_bits(nkey, old->key, bit, old_node_bit >> 1);
 		bit = (bit << 1) + 1; // assume comparisons with normal nodes
 
 		/* we must always check that our prefix is larger than the nodes
@@ -710,7 +711,7 @@
 			new->node.node_p = old->node.node_p;
 			up_ptr = &old->node.node_p;
 			new->node.bit = bit;
-			diff = cmp_bits(new->key, old->key, bit >> 1);
+			diff = cmp_bits(nkey, old->key, bit >> 1);
 			goto insert_above;
 		}
 
@@ -746,7 +747,7 @@
 		root = &old->node.branches;
 		side = old_node_bit & 7;
 		side ^= 7;
-		side = (new->key[old_node_bit >> 3] >> side) & 1;
+		side = (nkey[old_node_bit >> 3] >> side) & 1;
 		troot = root->b[side];
 	}
 
@@ -789,7 +790,7 @@
 	 */
 	diff = 0;
 	if (bit < old->node.pfx && bit < npfx)
-		diff = cmp_bits(new->key, old->key, bit);
+		diff = cmp_bits(nkey, old->key, bit);
 
 	if (diff == 0) {
 		/* Both keys match. Either it's a duplicate entry or we have to