v0.13.0 - C++ Acceleration Layer Improvements

• Extended C++20 kernel suite: added rolling_max, rolling_min via monotonic deque (O(n)), plus ATR kernel with Wilder smoothing
• Expanded numerical equivalence suite with per-kernel tolerance bounds verified to ≤ 4×10⁻⁹ vs pandas baseline
• Benchmark suite extended with multi-symbol parallel measurements (ThreadPoolExecutor) with 39.8× end-to-end on 8 × 1M rows
• CMake/scikit-build-core packaging improvements with cleaner pip install ./backend/cpp experience
• Graceful degradation confirmed: falls back to pandas silently when .so not built (Render production path)
• 152 tests passing