fft.h: include <cstring> for std::memcpy

[FakeSohi]

Summary

fft.h uses std::memcpy inside RealFFT<>::fft and RealFFT<>::ifft (lines 1074 and 1284) without including <cstring>. On older libstdc++ this header was pulled in transitively through other standard headers, so the build succeeded by accident. Newer toolchains (GCC 13+, which ships with MSYS2 UCRT64) have reduced the transitive-include graph, and the build now fails with:

fft.h: In member function 'void signalsmith::linear::RealFFT<Sample, splitComputation, halfBinShift>::fft(size_t, const Sample*, Complex*)':
fft.h:1074:46: error: 'memcpy' is not a member of 'std'; did you mean 'wmemcpy'?
 1074 |             std::memcpy(tmpTime.data(), time, sizeof(Complex)*hSize);
fft.h: In member function 'void signalsmith::linear::RealFFT<Sample, splitComputation, halfBinShift>::ifft(size_t, const Complex*, Sample*)':
fft.h:1284:46: error: 'memcpy' is not a member of 'std'; did you mean 'wmemcpy'?
 1284 |             std::memcpy(time, tmpTime.data(), sizeof(Complex)*hSize);

This one-line change adds an explicit #include <cstring> so the header is self-contained and compiles on any conformant toolchain, without affecting platforms that previously compiled by accident.

Reproduced while building signalsmith-stretch's C shim against MSYS2 UCRT64 (GCC 15.x):

g++ -O2 -std=c++17 -DNDEBUG -shared -static-libgcc -static-libstdc++ \
    -I.. -I. signalsmith_stretch_c.cpp -o signalsmith_stretch.dll

Test plan

• Rebuild the failing signalsmith-stretch command above under MSYS2 UCRT64 — compiles cleanly with the patch applied
• Reviewer: sanity-check on whatever CI the project uses (Linux/macOS/MSVC) to confirm no regression; the change is an additive #include so no behavior change is expected