Safe ConvolutionND

[bloodre]

Summary

This PR introduces safe convolution operations for dimension N>=4, related safe APIs, tests, and deprecates Conv2d operations. Work was started after reading this issue content.

Content

• Safe wrapper for cudnnSetFilterNdDescriptor
• Safe wrapper for cudnnSetConvolutionNdDescriptor
• Create of alias for Conv2dDescriptor, Conv2dForward, Conv2dBackwardData, Conv2dBackwardFilter: ConvDescriptor, ConvForward, ConvBackwardData, ConvBackwardFilter. The original struct names have also been deprecated. The reason is beside "2d" being included in the struct name, their methods apply to convolution operation in any dimension (pick_algorithm, get_workspace_size, launch).
• Addition of tests: one for 3D convolution (using new APIs) and another for 1D conv (sample code with Conv2d operation)

Regarding 1D convolution

As far as I understood from the CUDNN API reference, the ConvolutionNd are not used for 1D convolution:

• From cudnnSetTensorDescriptor: "Tensors are restricted to having at least 4 dimensions, and at most CUDNN_DIM_MAX dimensions (defined in cudnn.h). When working with lower dimensional data, it is recommended that the user create a 4D tensor, and set the size along unused dimensions to 1. ".
• cudnnConvolutionForward will return CUDNN_STATUS_BAD_PARAM if the input tensor descriptor and weight tensor description dimensions mismatch. Using 3-dimensional filters is therefore not possible for convolutions.
  As shown in the test "test_conv1d" that I added in src/cudnn/safe/mod.rs, a workaround is therefore to define a 4D tensor with the last dimension equal to 1, and to use 2D convolution operation with a flattened kernel.

Note

• It might be useful to create some utility function to generate the stride arrays of Nd tensors from the dimensions.

[coreylowman]

Looks good, thanks for this!