Add arbitrary-N real FFT via generalized Bruun (z^N-1 factorization)

[falseywinchnet]

Extend bfft.rfft/irfft to any N >= 2. Power-of-two stays on the native Bruun kernel; every other size uses a new generalized Bruun plan that factors z^N-1 over the reals -- the 2-adic part reuses the existing power-of-two SIMD cascade, and each odd prime factor peels off through a condition-1 real radix-p codelet (z^(pM)-1 = (z^M-1) * prod_{j}(z^(2M) - 2cos(2pij/p) z^M + 1)). FFT-grade for all N including primes (0.33-2.0x of NumPy/FFTW). Not Bluestein/Rader/mixed-radix.

• src/detail/genbruun_kernel.hpp: bruun::GenBruun (plan-time op schedule, owns scratch, forward + inverse; rooted pow2 cascades reuse norm_q_fwd/norm_q_inv)
• src/bfft.cpp: bfft_plan is now a tagged {RFFT | GenBruun}; bfft_plan_create accepts non-pow2; forward/inverse/size/bins dispatch; pow2-only entry points (native/magnitude/residue/f32) guarded -> BFFT_ERROR_INVALID_ARGUMENT
• bfft/_core.py, bfft/init.py: relax N check to N>=2; docstrings
• tests/genbruun_correctness.cpp (+ Makefile): make test now covers arbitrary N
• benchmarks/odd_prime_power_fftw_benchmark.cpp: FFTW-baseline benchmark, all N
• README "Arbitrary-N support"; notes/genbruun_port_handoff.md; src/detail/genbruun_acceleration_plan.md
• scratch_genbruun*.{cpp,py}: reference implementations

Perf: the arbitrary-N odd-radix projections are not yet SIMD, so non-pow2 sizes are ~3.5-5x slower than FFTW on prime powers (pow2 path unaffected). The acceleration plan documents the AVX2/NEON codelet work.