Fixes selection of cuDNN algorithm (pytorch#15881)

Summary:
This PR updates the logic for using cudnnGet* and cudnnFind*. Current version of cudnn find and get (v7) returns a pair of best algorithm and the convDesc mathType. While we were using the returned algorithm, we didn't update the mathType. As a result, we ended up with a slow choice of algorithm and math type. Without this patch, we are seeing a 10x regression in group convolutions.

Changelist:
- Changed the template arguments to be `perf_t` instead of `algo_t` to unify cudnnFind and cudnnGet. Both cudnnFind and cudnnGet have the same purpose and hence, it made sense to unify them and get rid of `getAlgorithm`.
- Used cudnnGet*_v7 everywhere cudnnGet* was being used.
- Removed all cudnn6 paths (This PR depends on pytorch#15851)

Differential Revision: D13787601

Pulled By: ezyang

fbshipit-source-id: 81fe86727673d021306fe1c99c3e528b7c9ad17f

aten/src/ATen/cudnn/Descriptors.h

@@ -155,9 +155,6 @@ struct AT_CUDA_API ConvolutionDescriptor
     AT_CUDNN_CHECK(cudnnSetConvolutionNdDescriptor(mut_desc(), dim, pad, stride, upscale,
                                           CUDNN_CROSS_CORRELATION, mathType));
     AT_CUDNN_CHECK(cudnnSetConvolutionGroupCount(mut_desc(), groups));
-    AT_CUDNN_CHECK(cudnnSetConvolutionMathType(mut_desc(), CUDNN_DEFAULT_MATH));
-    if(dataType == CUDNN_DATA_HALF)
-      AT_CUDNN_CHECK(cudnnSetConvolutionMathType(mut_desc(), CUDNN_TENSOR_OP_MATH));
   }
 };