ms_experimental DFT feedback

I have recently been using the DFT operator which is built with the ms_experimental flag and had some feedback and questions regarding the roadmap.

1. Performance: For input sizes which are power of 2 DFT dispatches to fft_radix2, else it falls back to dft_naive (which is incredibly slow). At a high level I understand why it is importance to have a power of 2 input for performance, but torch.fft.ftt is robust to other sizes and does not exhibit these same performance cliffs, so I was wondering if the onnxruntime implementation is able to be modified in the same way?
2. Experimental status: Are there plans to integrate these operators into the standard opset to avoid the need for the custom build flag in the future? There is an open PR to provide support for FFT as native ONNX operator, which could help fast-track this?
3. EP alignment: The ms_experimental op currently only supports the CPU execution provider, however there is a different fft implementation for CUDA in contrib_ops/cuda/math/fft_ops.h, which has a different function signature and domain, are there any plans to align these so that a single model could access both operators, just by switching out the EP?