Fix the large range issue on QuantizeAvx2 (pytorch#1784)

Summary:
Pull Request resolved: pytorch#1784

We have seg fault when the range of tensor quantization is big. The common int32_t vs. int64_t issue.

Reviewed By: sryap

Differential Revision: D46123138

fbshipit-source-id: d5f0c9c2b9a2d10159e7d227d1207dad3917aa87

src/QuantUtilsAvx2.cc

@@ -41,7 +41,7 @@ void QuantizeAvx2(
   // that is exactly representable in float
   constexpr int32_t int32_float_max_val =
       std::numeric_limits<int32_t>::max() - 127;
-  int i = 0;
+  int64_t i = 0;
   float inverse_scale = 1.f / qparams.scale;
   __m256 inverse_scale_v = _mm256_set1_ps(inverse_scale);
   // clang-format off