Correct sub pel offset checks for half pel point

In "aom_sub_pixel_avg_variance32xh_avx2" function, sub pel offset
is being checked with 8 to handle special cases of half pel
which is incorrect.

Above is modified to check for an offset of 4 (which is half pel).

BUG=aomedia:3007

Change-Id: I3150829989e701d582dbe025ac98f910c3394e55
diff --git a/aom_dsp/x86/variance_impl_avx2.c b/aom_dsp/x86/variance_impl_avx2.c
index f779270..163e4cc 100644
--- a/aom_dsp/x86/variance_impl_avx2.c
+++ b/aom_dsp/x86/variance_impl_avx2.c
@@ -616,7 +616,7 @@
         src += src_stride;
         dst += dst_stride;
       }
-    } else if (y_offset == 8) {
+    } else if (y_offset == 4) {
       __m256i src_next_reg;
       for (i = 0; i < height; i++) {
         LOAD_SRC_DST
@@ -652,8 +652,8 @@
         dst += dst_stride;
       }
     }
-    // x_offset = 8  and y_offset = 0
-  } else if (x_offset == 8) {
+    // x_offset = 4  and y_offset = 0
+  } else if (x_offset == 4) {
     if (y_offset == 0) {
       __m256i src_next_reg;
       for (i = 0; i < height; i++) {
@@ -668,8 +668,8 @@
         src += src_stride;
         dst += dst_stride;
       }
-      // x_offset = 8  and y_offset = 8
-    } else if (y_offset == 8) {
+      // x_offset = 4  and y_offset = 4
+    } else if (y_offset == 4) {
       __m256i src_next_reg, src_avg;
       // load source and another source starting from the next
       // following byte
@@ -691,7 +691,7 @@
         CALC_SUM_SSE_INSIDE_LOOP
         dst += dst_stride;
       }
-      // x_offset = 8  and y_offset = bilin interpolation
+      // x_offset = 4  and y_offset = bilin interpolation
     } else {
       __m256i filter, pw8, src_next_reg, src_avg;
       y_offset <<= 5;
@@ -741,8 +741,8 @@
         src += src_stride;
         dst += dst_stride;
       }
-      // x_offset = bilin interpolation and y_offset = 8
-    } else if (y_offset == 8) {
+      // x_offset = bilin interpolation and y_offset = 4
+    } else if (y_offset == 4) {
       __m256i filter, pw8, src_next_reg, src_pack;
       x_offset <<= 5;
       filter = _mm256_load_si256(