Fix tpl stats calculation

Tpl stats should be calculated on the basis of stats block size instead
of mi block size. This CL fixed the issue.

Borg test results at speed 1:
       avg_psnr:  ovr_psnr:  ssim:  vmaf:
hdres:  -0.012    -0.014    -0.014  -0.031
midres:  0.031     0.018     0.100  -0.066
lowres: -0.031    -0.035    -0.050  -0.158

STATS_CHANGED

Change-Id: If23018d20c34fbf5e382c54fc162369eaa10a6a0
diff --git a/av1/encoder/tpl_model.c b/av1/encoder/tpl_model.c
index 6a828a8..4890f37 100644
--- a/av1/encoder/tpl_model.c
+++ b/av1/encoder/tpl_model.c
@@ -697,13 +697,14 @@
   const int mi_height = mi_size_high[bsize];
   const int mi_width = mi_size_wide[bsize];
   const int step = 1 << block_mis_log2;
+  const int div = (mi_height >> block_mis_log2) * (mi_width >> block_mis_log2);
 
-  int64_t intra_cost = src_stats->intra_cost / (mi_height * mi_width);
-  int64_t inter_cost = src_stats->inter_cost / (mi_height * mi_width);
-  int64_t srcrf_dist = src_stats->srcrf_dist / (mi_height * mi_width);
-  int64_t recrf_dist = src_stats->recrf_dist / (mi_height * mi_width);
-  int64_t srcrf_rate = src_stats->srcrf_rate / (mi_height * mi_width);
-  int64_t recrf_rate = src_stats->recrf_rate / (mi_height * mi_width);
+  int64_t intra_cost = src_stats->intra_cost / div;
+  int64_t inter_cost = src_stats->inter_cost / div;
+  int64_t srcrf_dist = src_stats->srcrf_dist / div;
+  int64_t recrf_dist = src_stats->recrf_dist / div;
+  int64_t srcrf_rate = src_stats->srcrf_rate / div;
+  int64_t recrf_rate = src_stats->recrf_rate / div;
 
   intra_cost = AOMMAX(1, intra_cost);
   inter_cost = AOMMAX(1, inter_cost);