Add differentiable CPU backend for KDA research

[xy200303]

Summary

• add a differentiable recurrent KDA backend with a portable PyTorch reference path and an optional C++/ATen CPU operator
• provide analytical first-order CPU gradients, variable-length sequence handling, grouped value heads, FLA-style gates, and initial/final states
• support CPU-only and compiler-free installations while preserving the existing CUTLASS CUDA inference backend
• document the CPU backend as a research and prototyping path for integrating KDA into conventional PyTorch models and exploring model innovations

Verification

• FLASH_KDA_BUILD_CUDA=0 FLASH_KDA_BUILD_CPU=1 python setup.py build_ext --inplace
• python -m pytest tests/test_cpu.py -q with the compiled extension: 7 passed
• python -m pytest tests/test_cpu.py -q without the extension: 5 passed, 2 skipped
• git diff --check

Notes

• the intended use is research, model prototyping, architecture experimentation, correctness validation, and small workloads
• the C++ implementation currently composes ATen tensor operations; it is not a fused AVX/NEON kernel and does not claim production-grade long-sequence CPU throughput
• the analytical C++ backward supports first-order gradients; higher-order gradients can use use_cpp_backend=False
• Linux and the minimum supported PyTorch 2.4 still need CI coverage; current compiled validation used Windows and PyTorch 2.12.0+cpu
• native CUDA backward remains future work; the existing CUTLASS path remains the high-performance CUDA inference backend

Closes #26