PyTorch torch.fft / torch.stft fail with CUFFT_NOT_SUPPORTED on Windows + HIP 7.13 + RX 7900 XTX

RuntimeError: cuFFT error: CUFFT_NOT_SUPPORTED

### Steps to reproduce

## Summary

On Windows with current ZLUDA and HIP/ROCm 7.13, basic PyTorch CUDA operations work through ZLUDA, but all tested PyTorch FFT operations fail with:

RuntimeError: cuFFT error: CUFFT_NOT_SUPPORTED

This affects torch.fft.* and torch.stft, which blocks PyTorch audio workloads such as SpeechBrain / pyannote diarization.

Related older issue: #247

## System

OS: Windows
GPU: AMD Radeon RX 7900 XTX
GPU architecture: gfx110X
ZLUDA: recent Windows build
HIP/ROCm: 7.13, therock-dist-windows-gfx110X-all
Python: venv
PyTorch: 2.5.1+cu118
CUDA visible through ZLUDA: yes

## Environment setup

set "HIP713=K:\HIP713"
set "HIP_PATH=%HIP713%"
set "ROCM_PATH=%HIP713%"
set "PATH=%HIP713%\bin;%PATH%"

zluda\zluda.exe -- python zluda_fft_matrix_test.py

DLL resolution:

K:\HIP713\bin\rocblas.dll
K:\HIP713\bin\MIOpen.dll
K:\HIP713\bin\amdhip64_7.dll

## What works

The following PyTorch CUDA operations work through ZLUDA:

import torch
import torch.nn as nn

print(torch.__version__)
print(torch.cuda.is_available())
print(torch.cuda.get_device_name(0))

x = torch.randn(2048, 2048, device="cuda", dtype=torch.float32)
y = x @ x
print(y.mean().item())

m = nn.Conv1d(80, 256, kernel_size=3).cuda().float()
x = torch.randn(1, 80, 1000, device="cuda", dtype=torch.float32)
print(m(x).shape)

torch.backends.cudnn.enabled = False
m = nn.LSTM(80, 256, 2).cuda().float()
x = torch.randn(100, 1, 80, device="cuda", dtype=torch.float32)
print(m(x)[0].shape)

Observed result:

torch.cuda.is_available() = True
device = AMD Radeon RX 7900 XTX [ZLUDA]
matmul works
Conv1d works
LSTM works with cuDNN disabled

## What fails

All tested CUDA FFT variants fail:

import torch

x = torch.randn(1024, device="cuda", dtype=torch.float32)

torch.fft.fft(x.to(torch.complex64))
torch.fft.ifft(x.to(torch.complex64))
torch.fft.rfft(x)
torch.fft.irfft(torch.randn(513, device="cuda").to(torch.complex64), n=1024)
torch.fft.fft2(torch.randn(64, 64, device="cuda").to(torch.complex64))
torch.stft(x, n_fft=400, hop_length=160, win_length=400, return_complex=True)

Each one fails with:

RuntimeError: cuFFT error: CUFFT_NOT_SUPPORTED

Tested FFT sizes:

16
64
256
512
1024
2048
4096
16000

The CPU reference works:

import torch

x = torch.randn(16000, device="cpu", dtype=torch.float32)
window = torch.hann_window(400, device="cpu", dtype=torch.float32)

y = torch.stft(
    x,
    n_fft=400,
    hop_length=160,
    win_length=400,
    window=window,
    return_complex=True,
)

print(y.shape)

Observed result:

CPU-STFT works
CUDA FFT/STFT fails with CUFFT_NOT_SUPPORTED

## Real workload impact

This blocks audio workloads that use PyTorch FFT/STFT on CUDA, for example:

pyannote
SpeechBrain
speaker diarization pipelines
torch.stft based feature extraction

In a real SoniTranslate / WhisperX / pyannote workflow, the crash occurs here:

pyannote
-> SpeechBrain
-> compute_features
-> torch.stft
-> RuntimeError: cuFFT error: CUFFT_NOT_SUPPORTED

## Question

Is cuFFT support currently expected to be unavailable in ZLUDA on Windows?

Is CUFFT_NOT_SUPPORTED expected because zluda_fft is currently incomplete/stubbed, or is this a bug in the current Windows + HIP 7.13 path?

I can provide the full test script and logs if useful.

### ZLUDA version

Version 6-preview.69

### Operating System

Win11 2025H2 26200.8328

### GPU

AMD Radeon RX 7900XTX