tscc2: do not add/subtract 128 bias during DCT

It turns out that the reference decoder subtracts 128 from DC during block
decode but adds it back during reordering block with zigzag pattern.
Transforming block with incorrect DC caused heavy visual artifacts for
many quantisers.

libavcodec/tscc2.c

@@ -95,7 +95,7 @@ static av_cold int init_vlcs(TSCC2Context *c)
     OP(d3, 5 * ((s0) - (s1) + (s2)) - 2 * (s3));  \
 
 #define COL_OP(a, b)  a = b
-#define ROW_OP(a, b)  a = (((b) + 0x20) >> 6) + 0x80
+#define ROW_OP(a, b)  a = ((b) + 0x20) >> 6
 
 static void tscc2_idct4_put(int *in, int q[3], uint8_t *dst, int stride)
 {
@@ -158,7 +158,7 @@ static int tscc2_decode_mb(TSCC2Context *c, int *q, int vlc_set,
             }
             dc          = (dc + prev_dc) & 0xFF;
             prev_dc     = dc;
-            c->block[0] = dc - 0x80;
+            c->block[0] = dc;
 
             nc = get_vlc2(gb, c->nc_vlc[vlc_set].table, 9, 1);
             if (nc == -1)