Fix GPU segfaults in PW USPP calculations

[chengleizheng]

Summary

Fix GPU segfaults in PW USPP calculations caused by CPU code accessing GPU
memory pointers. Added CPU intermediate buffers and D2H/H2D sync patterns
in three USPP-specific code paths.

Changes

• hamilt_pw: sPsi reads qq_nt from CPU backup (qq_nt.ptr) instead of
  GPU pointer, fills a CPU buffer, then syncs to GPU for cuBLAS.

• elecstate_pw (cal_becsum): becp computed on GPU via gemm then synced
  to CPU; auxk1/auxk2 use CPU vectors; aux_gk computed on GPU then synced
  to CPU; becsum allocated on CPU (consumed by CPU-only force/stress code).

• elecstate_pw (add_usrho): Added D2H/H2D wrappers around real2recip
  and recip2real FFT calls, which read/write via CPU loops.

• elecstate_pw (addusdens_g): Rewritten to use pure CPU computation
  (CPU BLAS zgemm_, CPU radial_fft_q); ylmk0 computed on GPU with H2D'd
  gcar then D2H'd back.

Test plan

• Run PW USPP test cases (003, 008, 009) on GPU, verify no segfault
• Run compute-sanitizer to confirm 0 GPU memory errors

[chengleizheng]

Benchmark: Fe atom single-point — device=gpu

Hardware: i9-14900HX + RTX 4060 (x1), 32 threads
System: 1 Fe atom, USPP, 9×9×9 grid, collinear spin, 8 k-points

SCF convergence

Iter	TMAG	AMAG	ETOT/eV	EDIFF/eV	DRHO	TIME/s
CG1	1.25e+00	1.25e+00	-2860.20876	0.00e+00	4.22e+01	1.38
CG2	7.50e-01	7.50e-01	-2715.34625	1.45e+02	1.42e+01	0.87
CG3	9.52e-01	9.52e-01	-2736.26617	-2.09e+01	3.15e-01	0.94
CG4	8.94e-01	8.94e-01	-2736.93124	-6.65e-01	2.63e-01	0.78
CG5	7.51e-01	7.51e-01	-2737.65342	-7.22e-01	5.83e-02	0.82
CG6	7.50e-01	7.50e-01	-2738.69933	-1.05e+00	3.02e-02	0.79
CG7	7.50e-01	7.50e-01	-2738.89917	-2.00e-01	1.72e-02	0.95
CG8	7.50e-01	7.50e-01	-2739.30581	-4.07e-01	1.11e-02	0.91
CG9	7.49e-01	7.49e-01	-2739.33261	-2.68e-02	9.13e-04	1.04
CG10	7.48e-01	7.48e-01	-2739.33515	-2.54e-03	8.27e-05	1.01
CG11	7.46e-01	7.46e-01	-2739.33520	-5.26e-05	8.06e-05	0.95
CG12	7.47e-01	7.47e-01	-2739.33531	-1.02e-04	1.11e-06	1.07
CG13	7.47e-01	7.47e-01	-2739.33531	-4.71e-06	4.53e-07	1.07
CG14	7.46e-01	7.46e-01	-2739.33531	-1.28e-06	4.57e-08	1.12
CG15	7.46e-01	7.46e-01	-2739.33531	-3.13e-07	1.75e-09	1.16
CG16	7.46e-01	7.46e-01	-2739.33531	-1.20e-07	4.32e-11	1.33

Final ETOT: -2739.33531 eV

Stress (kbar)

     -92851.610       -0.0073       -0.0056
         -0.0073   -92851.612       -0.0011
         -0.0056       -0.0011   -92851.609

TOTAL-PRESSURE: -92851.610 kbar

Timing (total 17.22s)

Hotspot	Time/s	Pct
diag_once (CG)	10.91	63.4%
pp_cell_vnl::getvnl	1.95	11.3%
diag_subspace	2.49	14.5%
Operator::hPsi	1.85	10.8%
diag_hegvd	1.32	7.7%
Operator::veff_pw	1.38	8.0%
ElecStatePW::psiToRho	0.99	5.7%
PW_Basis_K::recip_to_real gpu	0.96	5.6%

Benchmark: Fe atom single-point — device=cpu

Hardware: i9-14900HX, 32 threads
System: 1 Fe atom, USPP, 9×9×9 grid, collinear spin, 8 k-points

SCF convergence

Iter	TMAG	AMAG	ETOT/eV	EDIFF/eV	DRHO	TIME/s
CG1	1.25e+00	1.25e+00	-2860.20876	0.00e+00	4.22e+01	0.52
CG2	7.50e-01	7.50e-01	-2715.34625	1.45e+02	1.42e+01	0.76
CG3	9.52e-01	9.52e-01	-2736.26617	-2.09e+01	3.15e-01	1.51
CG4	8.94e-01	8.94e-01	-2736.93124	-6.65e-01	2.63e-01	1.17
CG5	7.51e-01	7.51e-01	-2737.65342	-7.22e-01	5.83e-02	0.76
CG6	7.50e-01	7.50e-01	-2738.69933	-1.05e+00	3.02e-02	0.65
CG7	7.50e-01	7.50e-01	-2738.89917	-2.00e-01	1.72e-02	0.89
CG8	7.50e-01	7.50e-01	-2739.30581	-4.07e-01	1.11e-02	0.47
CG9	7.49e-01	7.49e-01	-2739.33261	-2.68e-02	9.13e-04	0.82
CG10	7.48e-01	7.48e-01	-2739.33515	-2.54e-03	8.27e-05	0.41
CG11	7.46e-01	7.46e-01	-2739.33520	-5.25e-05	8.06e-05	1.16
CG12	7.47e-01	7.47e-01	-2739.33531	-1.02e-04	1.11e-06	0.39
CG13	7.47e-01	7.47e-01	-2739.33531	-4.71e-06	4.53e-07	0.72
CG14	7.46e-01	7.46e-01	-2739.33531	-1.28e-06	4.57e-08	0.64
CG15	7.46e-01	7.46e-01	-2739.33531	-3.15e-07	1.75e-09	0.83
CG16	7.46e-01	7.46e-01	-2739.33531	-1.20e-07	4.32e-11	0.97

Final ETOT: -2739.33531 eV

Stress (kbar)

     -92851.610       -0.0073       -0.0056
         -0.0073   -92851.612       -0.0011
         -0.0056       -0.0011   -92851.609

TOTAL-PRESSURE: -92851.610 kbar

Timing (total 14.01s)

Hotspot	Time/s	Pct
diag_once (CG)	9.19	65.6%
Operator::veff_pw	4.03	28.8%
PW_Basis_K::recip2real	2.92	20.9%
diag_subspace	2.69	19.2%
diag_hegvd	1.93	13.8%
PW_Basis_K::real2recip	1.18	8.4%
PSIPrepare::initialize_psi	0.94	6.7%
pp_cell_vnl::getvnl	0.72	5.1%
Nonlocal::getvnl	0.69	4.9%
ElecStatePW::psiToRho	0.50	3.6%

[mohanchen]

Great job figuring out this GPU bug with your efforts, and thanks a lot for your contribution.