Avoid lrint() in speed detection (rounding doesn't improve results).

Signed-off-by: Stefan Westerfeld <stefan@space.twc.de>

src/TODO

 speed detection possible improvements:
 - refine with 128 sync step instead of 256?
 - grid size adjustment
-- remove lrint?
 - detect silence instead of total volume?
 - split big region into two smaller ones?
 - connected search on even smaller regions?

src/wmspeed.cc

@@ -302,14 +302,12 @@ SpeedSync::compare_bits (BitValue *bit_values, double relative_speed, int *start
 
   for (auto si = sync_bits.begin() + mi; si != sync_bits.end(); si++)
     {
-      int index = offset + si->frame * steps_per_frame;
-
-      const int index1 = lrint (index * relative_speed_inv);
-      if (index1 >= sync_matrix.rows())
+      int index = (offset + si->frame * steps_per_frame) * relative_speed_inv;
+      if (index >= sync_matrix.rows())
         return;
 
       auto& bv = bit_values[si->bit];
-      auto mags = sync_matrix (index1, mi);
+      auto mags = sync_matrix (index, mi);
       if (B2)
         {
           bv.umag += mags.dmag;