[Pytorch] enable mixed precision fp16 training.

* enabled miopen fp16 for mixed precision training
* removed ununsed mathType completely

Author: lcskrishna

aten/src/ATen/miopen/Descriptors.h

@@ -122,8 +122,6 @@ struct ConvolutionDescriptor
                       &miopenDestroyConvolutionDescriptor>
 {
   void set(miopenDataType_t dataType, int dim, int* pad, int* stride, int * upscale /* aka dilation */, int groups) {
-    miopenDataType_t mathType = dataType;
-    if (dataType == miopenHalf) mathType = miopenFloat;
     MIOPEN_CHECK(miopenInitConvolutionDescriptor(mut_desc(), miopenConvolution, pad[0], pad[1], stride[0], stride[1], upscale[0], upscale[1]));
     MIOPEN_CHECK(miopenSetConvolutionGroupCount(mut_desc(), groups));
   }

aten/src/ATen/native/Normalization.cpp

@@ -265,7 +265,7 @@ Tensor batch_norm(
   bool use_miopen = (input.is_cuda()
                && input.dim() <= MIOPEN_DIM_MAX
                && input.type().scalarType() != at::kDouble
-               && (input.type().scalarType() == weight.type().scalarType())
+               && (weight.type().scalarType() != at::kHalf)
                && weight.defined() && bias.defined()
                && ((running_mean.defined() && running_var.defined())
                  || (!running_mean.defined() && !running_var.defined() && training))