2021.2

• Enable transforms for any array size (VkFFT Bluestein algorithm)
• Allow 3D transforms on arrays with >3 dimensions (batch transform)
• Support for transforms on a given list of axes, instead of
  only the first ndim axes. Unavailable for R2C.
• Added a simple pyvkfft.fft interface with fftn, ifftn, rfftn, irfftn
  functions which automatically recognize the type of GPU arrays
  and cache the generated VkFFTApp (FFT plans).
• Direct Cosine Transform (DCT) of type 2, 3 and 4 (EXPERIMENTAL)
• Support CuPy arrays in addition to PyCUDA and PyOpenCL
• OpenCL: test for half and double-precision support
• OpenCL: relax accuracy requirements in unit tests
• Fix shape test for out-of-place R2C transforms
• Add a base VkFFTApp class common to OpenCL and CUDA
• Installation: fix macOS compilation. Allow selection of backends
  from an environment variable