﻿Implement GPU FFT via cuFFT for spectrum pipeline (Windows). Add optional build tag `cufft` and fall back to CPU FFT if not available. Requirements:
- Add CGo bindings to cuFFT (cufftPlan1d, cufftExecC2C) for complex64 input.
- Use CUDA runtime to allocate device buffers and transfer IQ samples; keep a simple GPU FFT cache (plan reuse).
- Add config flag `use_gpu_fft` (bool) and UI toggle. If enabled and GPU available, use cuFFT; otherwise fall back.
- Keep existing CPU path intact.
- Update README with CUDA prerequisites (CUDA toolkit, PATH, CGO flags).
- Add a tiny diagnostic endpoint `/api/gpu` that reports whether GPU FFT is active.
- Build & run: go build -tags "sdrplay,cufft" ./cmd/sdrd.
- Run tests (go test ./...).
- Commit changes.
