[Feature Request] Add DFT support for CUDAExecutionProvider

[alexander-veselov]

Describe the feature request

While running ONNX model that contains DFT operation using CUDAExecutionProvider, I see messages like: "CUDA kernel not found in registries for Op type: DFT".

Is DFT supported while using CUDAExecutionProvider? Or is this something with my model?
DFT is performance bottleneck for my model, so it would be really nice to have CUDA support for DFT operation.

There is NVidia cuFFT library. Seems to be exactly what I need.

1. Can I expect that DFT operation will be implemented for CUDAExecutionProvider in near future?
2. Can I workaround my problem somehow to run DFT on GPU?

Describe scenario use case

Better performance for DFT operation when CUDAExecutionProvider is available.