cc @cudawarped @chacha21. I want to drop gridTransformTuple and some hacky things on top of tuple, because it's hard to maintain it with different CUDA versions. What do you think about the solution and further cleanup?

@asmorkalov Unfortunatley it removes the fusion as the source has to be read from global memory twice but I can't think of a less involved fix. The alternative of adding two ops to gridTransformUnary as you did with gridTransformBinary is probably overkill as we should probably be moving away from cudev and towards Thrust for these operations if the CUDA namespace has a future.

@cudawarped Thanks for the opinion.

Another solution would be to drop the cudev/functional abstraction for polarToCart/cartToPolar and use dedicated CUDA kernels to handle the different cases