fix(gpu): FFT_Bundle host-side FFT fallback to fix Wannier90 segfault on GPU

[chengleizheng]

Fix #7770

This is the follow-up announced in #7847, which fixed the LDOS and wfc-read crashes of this issue; this PR fixes the remaining one, the Wannier90 GPU segfault (101_PW_W90, 100_PW_W90 in the issue).

Root cause. In a GPU build only the wavefunction basis (pw_wfc) uses a GPU FFT backend; all rho-family bases always keep a CPU backend. The Wannier90 output runs on the host CPU mirror (stp.sync_cpu() in ctrl_output_pw.cpp) and performs its transforms through the host-memory FFT interface (recip2real/real2recip -> fftxyfor/fftzbac/get_auxg_data/...). The GPU backends (FFT_CUDA/FFT_ROCM) implement only the device-memory 3D-FFT interface; the host-memory virtuals are inherited from FFT_BASE, where they are declared __attribute__((weak)) without a definition, so on ELF their vtable slots are null. The first call of this kind (e.g. get_auxg_data() in PW_Basis_K::recip2real, pw_transform_k.cpp) jumps to address 0 -- the null-pointer segfault reported in the issue.

Fix. When device != "cpu", FFT_Bundle::initfft now creates an additional FFT_CPU backend per active precision (host_fft_float/host_fft_double) that serves the host-memory interface, and all host-memory entry points are routed through get_host_fft_float()/get_host_fft_double(). The device-memory 3D-FFT path (fft3D_forward/backward, get_auxr_3d_data) still uses the device backend, and when device == "cpu" the selectors return the original backends, so CPU builds are unchanged (no extra object is created). The same fallback also covers the identical latent crash in the Berry-phase output (unk_overlap_pw), the ML-KEDF descriptors and DFPT, which all use the host-memory interface of the wfc bundle.

We kept the host fallback instead of routing these calls through the recip2real_gpu device path used in #7847 because these consumers work on host-resident arrays and hold no device copy of the spectra: the device path would add H2D/D2H traffic and buffer-layout work, and it currently requires gamma_only == false and poolnproc == 1. Serving the host interface from an FFT_CPU instance -- exactly what a CPU bundle has always done -- is the minimal fix; the only cost is one extra host FFTW plan set (plus its two work buffers) per wavefunction basis in GPU builds.

Verification:

• CUDA build and CPU build of this branch compile; CPU results unchanged.
• GPU: 101_PW_W90 previously segfaulted during the Wannier90 output; it now completes. Against the CPU reference, eigenvalues agree to ~1e-12 eV, |amn| to 5.7e-8, |mmn| exactly. amn/mmn differ only by the intrinsic per-eigenstate gauge phases.
• 101_PW_W90 stays disabled in CASES_GPU.txt: catch_properties.sh compares amn/mmn element-wise (Re/Im), which cannot accept a gauge rotation; re-enabling the case needs a phase-invariant comparison for these two files as a follow-up.

No INPUT or docs change; the device 3D-FFT path is untouched.

[chengleizheng]

before the fix, when running the case, we get:

[ RUN      ] 101_PW_W90/
[LAPTOP-G5PS9KT3:42553] *** Process received signal ***
[LAPTOP-G5PS9KT3:42553] Signal: Segmentation fault (11)
[LAPTOP-G5PS9KT3:42553] Signal code: Address not mapped (1)
[LAPTOP-G5PS9KT3:42553] Failing at address: (nil)
[LAPTOP-G5PS9KT3:42553] [ 0] /lib/x86_64-linux-gnu/libc.so.6(+0x42520)[0x76ad36842520]
[LAPTOP-G5PS9KT3:42553] *** End of error message ***
--------------------------------------------------------------------------
Primary job  terminated normally, but 1 process returned
a non-zero exit code. Per user-direction, the job has been aborted.
--------------------------------------------------------------------------
--------------------------------------------------------------------------
mpirun noticed that process rank 0 with PID 0 on node LAPTOP-G5PS9KT3 exited on signal 11 (Segmentation fault).
......

after the fix, we have the result:

[ RUN      ] 101_PW_W90/
Compare Error: line 2, column 4
 diamond.amn: -0.538560137556
 OUT.autotest/diamond.amn: -0.768120379683
 [ERROR     ]  Fatal Error in catch_properties.sh
[----------] Time elapsed: 13.004 seconds

../integrate/Autotest.sh: line 294: /tmp/abacus-autotest.hVQpii/101_PW_W90/.res: No such file or directory
[ PASSED   ]  0 test cases passed.

Running the GPU version, ETOT = -308.0061503011086 vs -308.0061503011041 in result.ref.
This PR fixes the GPU crash. The remaining W90 ERROR is an expected behaviour(gauge freedom in amn/mmn), not a bug; enabling the test on GPU would need a gauge-invariant comparison.

[ABACUS-PR-BOT]

Static review of 8e4299c08dcff5237b7010ebf4e5ba7a85e5499e, independently checked by a second reviewer. The remaining single-precision output gap and a regression-test suggestion are inline.

Non-blocking: please add the exact CPU/CUDA configure/build, 101_PW_W90 run and comparison commands used for the reported results, including precision, ENABLE_FLOAT_FFTW, rank count and executable revision, so the manual verification can be reproduced.

git diff --check passed. The pinned governance check exited 0; its warnings were reviewed individually. No builds or runtime tests were run under this static-review request. CI diagnostics were unavailable through the connector (sign-in required).

[chengleizheng]

At the suggestion of ABACUS-PR-BOT, host_fft_double is now created whenever device != "cpu", independently of double_flag, and its setupFFT() / cleanFFT() / clear() calls are null-guarded instead of being nested under if (double_flag).