Add radix-4 RFFT kernel and remove duplicate dfft subtree

[falseywinchnet]

Motivation

• Remove a duplicated dfft/ workspace from the repository to keep the tree focused and small.
• Convert the provided standalone radix-4 RFFT prototype into a reusable internal kernel that fits the project's SIMD abstraction and heap-backed planning model.
• Provide both scalar and two-lane SIMD entry points so the kernel can be exercised as a drop-in internal primitive and iterated for backend-specific optimizations.

Description

• Deleted the large dfft/ duplicate subtree and added src/detail/radix4_rfft_kernel.hpp implementing a bruun::radix4_rfft_kernel class that owns planning tables (bit-reversal, cos/sin, staged twiddles) and exposes forward_scalar and forward_simd entry points.
• Implemented the radix-4 algorithm pieces from the prototype: bit-reverse input packing, fused first binomial, merge2 and merge4_inplace passes, scalar and SIMD butterfly code paths, and final packing to real/imag output bins.
• Integrated with the shared Bruun SIMD primitives and heap_array allocator so the kernel selects the existing two-lane SIMD backend where available and falls back to scalar code otherwise.

Testing

• Built and ran a small smoke driver that exercises radix4_rfft_kernel::forward_scalar with N=64 and validates output against a direct DFT, which completed successfully.
• Ran the project test target (make test), which executed the suite and reported correctness on the build host; the test run succeeded.
• Also compiled and ran a quick SIMD smoke compile for the new kernel and observed expected numeric results for multiple sizes during validation.

---

Codex Task