Convolution improvements

[marton78]

Convolution improvements — const setup, unscaled zconvolve, zero-phase convolution, _USE_MATH_DEFINES guard

Implements all four suggestions from #101 for both the float and double variants (which share one implementation header, so every change lands in both automatically). One suggestion is implemented with a design refinement specific to PFFFT's memory layout — see Zero-phase convolution below.

1. const setup pointers

All functions that only consume the setup (pffft_transform, pffft_transform_ordered, pffft_zreorder, pffft_zconvolve_*) now take const PFFFT_Setup * / const PFFFTD_Setup *. Verified read-only: the twiddle/ifac tables are written exclusively by the *_ti1_ps() init functions at setup creation; rfftf1_ps/cfftf1_ps already took (const float *wa, const int *ifac). Source-compatible with every existing caller; makes sharing one setup across threads formally correct.

2. Unscaled pffft_zconvolve()

New entry point computing dft_ab = dft_a * dft_b with no accumulation and no scaling argument. The scaling factor was a runtime value in the previous kernels, so the compiler could not fold away the dead multiply; removing it saves one vector multiply per SIMD group. As suggested in the issue, rescaling belongs on the static filter side.

Naming: the fork-specific pffft_zconvolve_no_accu() (which upstream never had) is renamed to pffft_zconvolve_scale() so the family reads by what each function does:

pffft_zconvolve_accumulate(setup, a, b, ab, scaling); /* ab += a*b*scaling */
pffft_zconvolve_scale     (setup, a, b, ab, scaling); /* ab  = a*b*scaling */
pffft_zconvolve           (setup, a, b, ab);          /* ab  = a*b         */

The old name remains as an exported deprecated synonym forwarding to the new symbol — existing binaries keep linking.

3. Zero-phase convolution — pffft_zconvert_zp() + pffft_zconvolve_zp()

Implements the r8brain trick for linear-phase FIRs centered at t=0 with left-hand taps wrapped around the block. One necessary refinement to the proposal: in PFFFT's unordered output the coefficients are not stored as interleaved re/im pairs — whole SIMD vectors alternate between real parts and imaginary parts, and F(0)/F(N/2) pack into the first lanes of vectors 0/1 (with the real-valued Nyquist coefficient sitting in an otherwise-imaginary lane). A flat sequential multiply therefore cannot work on unordered buffers directly. Instead this PR provides:

• pffft_zconvert_zp(setup, in, out, scaling) — converts an unordered REAL forward spectrum once at filter-setup time: imaginary vector lanes are zeroed, real lanes scaled, the DC/Nyquist pair preserved component-wise.
• pffft_zconvolve_zp(setup, x, hzp, ab) — multiplies any number of block spectra against the converted filter using two vector multiplies per re/im vector pair where a full complex multiply needs four (Y_re = X_re·H_re, Y_im = X_im·H_re).

This keeps the faster unordered transforms end-to-end, which is the point of the optimization. Measured vs zconvolve_scale(.., 1.0) on M2/NEON, best of 5: 1.53× / 1.60× / 1.59× / 1.47× at N = 128/512/2048/8192. Both functions return nonzero for non-REAL setups; aliasing (in == out) is supported and documented, including the DC/Nyquist lane caveat in the public headers.

C++ wrapper: convolve(a, b, ab), zconvertZP() and convolveZP() added to the public Fft<T> template (raw-pointer and AlignedVector forms); COMPLEX instantiations forward unchanged and are rejected by the underlying functions.

4. _USE_MATH_DEFINES

Now guarded with #ifndef so defining it as a compiler-level option no longer conflicts.

Also included

• Pre-existing bug fixes surfaced by the new tests: pffft_zconvolve_scale_nosimd() accumulated instead of assigned at the two REAL boundary lanes (bins 0 and N/2), and test_fft_factors looped forever under SIMD_SZ == 1 because min_fft_size(COMPLEX) returns 1 there (N += N_min/2 stepped by zero).
• CI: new build_no-simd_novect configuration (PFFFT_USE_SIMD=OFF -DPFFFT_USE_SCALAR_VECT=OFF) that compiles and ctests the SIMD_SZ==1 branch — previously never executed anywhere.

Testing

• New ctest coverage: bit-exact agreement between zconvolve and zconvolve_scale(.., 1.0); zero-phase round-trip against independently computed time-domain circular convolution; sentinel-based invariant that conversion writes every destination lane; scaling linearity; in-place aliasing equivalence; REAL-only guards; wrapper-level equivalents of all of the above. Red→green verified for the lane-zeroing fix.
• Full suite green in four configurations: default SIMD, PFFFT_USE_SIMD=OFF (4×scalar emulation), PFFFT_USE_SIMD=OFF + PFFFT_USE_SCALAR_VECT=OFF (SZ=1 nosimd), and float-less double-only builds.