[Feature] Faster hologram reconstruction
========================================

 * Added :class:`~libertem_holo.base.plan.ReconstructionPlan`, which
   precomputes once which part of a hologram's spectrum ends up in the output
   window, and then reconstructs each hologram with a real-to-complex FFT,
   transforming only the needed columns of the spectrum and gathering the
   window directly, without full-size :code:`roll`, :code:`fftshift` or
   scaling passes. Build it with
   :meth:`~libertem_holo.base.plan.ReconstructionPlan.from_params` and reuse
   it for a whole stack. Reconstructing the wave and the brightfield image of
   a 4096x4096 hologram at 1024x1024 takes about a sixth of the time it did.
 * :func:`~libertem_holo.base.reconstr.reconstruct_frame`,
   :func:`~libertem_holo.base.reconstr.reconstruct_bf` and
   :class:`~libertem_holo.base.reconstr.HologramFFT` are now thin wrappers
   around the plan, with the same signatures; results agree with the previous
   implementation up to floating point rounding.
   :func:`~libertem_holo.base.reconstr.reconstruct_bf` gained the
   :code:`precision` parameter. The index arrays of the plan are cached per
   geometry and array module, on the GPU for cupy, so calling the wrappers
   once per frame does not transfer anything per call. An :code:`out_shape`
   larger than the hologram, or a :code:`slice_fft` that is not the centered
   slice of :func:`~libertem_holo.base.utils.get_slice_fft`, now raise
   :code:`ValueError` instead of producing wrapped-around results.
 * The :code:`precision` parameter now selects the computation dtype
   explicitly: :code:`True` transforms the hologram in float64 (complex128
   results), :code:`False` in float32 (complex64 results). Previously, float32
   holograms were transformed in float32 regardless, and
   :code:`precision=False` still returned complex128.
 * numpy >= 2.0 is now required.