Pr501 fft final

[aaadelmann]

## Summary

This PR ports the FFT backend refactor from the larger PR501 branch onto current `master`, keeping the scope limited and reviewable.

The main change is to split the former monolithic FFT implementation into smaller backend and transform modules while preserving the existing public FFT usage for the current HeFFTe-backed transforms.

## What Is Included

### FFT Core Refactor

The FFT implementation is now organized into separate headers:

- `src/FFT/Traits.h`
- `src/FFT/Backend/Backend.h`
- `src/FFT/Backend/Heffte.h`
- `src/FFT/Transform/Common.h`
- `src/FFT/Transform/CC.h`
- `src/FFT/Transform/RC.h`
- `src/FFT/Transform/Trig.h`
- `src/FFT/Transform/Transform.h`

`src/FFT/FFT.h` is now the aggregate include for the refactored FFT module.

The old monolithic implementation header:

- `src/FFT/FFT.hpp`

was removed.

### Pruned Transforms

This PR also adds the pruned transform specializations:

- `src/FFT/Transform/PrunedCC.h`
- `src/FFT/Transform/PrunedRC.h`

These are included through the FFT transform aggregate header.

### Optional cuFFTMp Backend

This PR adds the multi-node NVIDIA cuFFTMp backend as an explicitly opt-in backend:

- `src/FFT/Backend/CuFFTMp.h`

A new CMake option was added:

```cmake
-DIPPL_ENABLE_CUFFTMP=ON

The option is guarded so that cuFFTMp can only be enabled when FFT and CUDA are both enabled:

• IPPL_ENABLE_CUFFTMP=ON requires IPPL_ENABLE_FFT=ON
• IPPL_ENABLE_CUFFTMP=ON requires CUDA in IPPL_PLATFORMS

The backend header is included only when both conditions are active:

#if defined(IPPL_ENABLE_CUFFTMP) && defined(KOKKOS_ENABLE_CUDA)
#include "FFT/Backend/CuFFTMp.h"
#endif

Plain single-node CuFFT.h support is intentionally not included in this PR.

What Is Intentionally Not Included

The original pr501-fft branch also contained changes to FFTOpenPoissonSolver and solver tests. Those were reviewed but not ported.

Reason: current master already contains newer Poisson solver functionality, including integrated and shifted Green-function support. The corresponding hunks from pr501-fft would remove that current coverage and regress master.

Specifically, this PR does not remove:

• GreenFunction::{STANDARD, INTEGRATED}
• integrated Green-function support
• shifted integrated Green-function support
• TestGaussianIntegrated
• TestIntegratedGreensFunction
• current CG, preconditioner, and multigrid tests

Validation

Configured and built an OpenMP FFT/unit-test build:

cmake -S . -B build-pr501-fft-final-openmp \
  -DIPPL_PLATFORMS=OPENMP \
  -DIPPL_ENABLE_FFT=ON \
  -DIPPL_ENABLE_UNIT_TESTS=ON \
  -DIPPL_ENABLE_TESTS=ON

cmake --build build-pr501-fft-final-openmp -j 8
ctest --test-dir build-pr501-fft-final-openmp --output-on-failure

Result:

100% tests passed, 0 tests failed out of 70

Configured and built a solver-enabled OpenMP build:

cmake -S . -B build-pr501-fft-final-solvers-openmp \
  -DIPPL_PLATFORMS=OPENMP \
  -DIPPL_ENABLE_FFT=ON \
  -DIPPL_ENABLE_SOLVERS=ON \
  -DIPPL_ENABLE_UNIT_TESTS=ON \
  -DIPPL_ENABLE_TESTS=ON

cmake --build build-pr501-fft-final-solvers-openmp -j 8
ctest --test-dir build-pr501-fft-final-solvers-openmp --output-on-failure

Result:

100% tests passed, 0 tests failed out of 85

Also verified the cuFFTMp CMake guard:

-DIPPL_PLATFORMS=OPENMP -DIPPL_ENABLE_FFT=ON -DIPPL_ENABLE_CUFFTMP=ON

fails clearly with:

IPPL_ENABLE_CUFFTMP requires CUDA in IPPL_PLATFORMS

Scaling

Notes

CUDA/cuFFTMp backend code is included behind compile-time guards and was not fully exercised locally because the local build is OpenMP-only. The default path remains HeFFTe, and non-CUDA builds do not include the cuFFTMp backend header.

[srikrrish]

I just noticed that apart from the PCG PR Paul is not mentioned as author in the previous PR (HOSG final) as well as this one. This needs to be fixed. Since HOSG final is already merged it may be difficult to correct it there and not sure (sorry should have checked more carefully). but hopefully at least in this PR and in the future ones it can be corrected.

[aaadelmann]

Will add Paul to this one and see if we can add him to the other as well.

[srikrrish]

It was decided that Paul will be added as a co-author at the end of all the PRs from 501 by selecting the individual commits. Hence, I am approving now.