Restore GPU-backed UInt128 pow2 Montgomery helper

todo.txt

@@ -1,2 +1,3 @@
-- Implement GpuRational struct to extend GpuUInt128 range. Avoid extending method calls stack. Instead prefer code duplication for a better performance. Create unit tests for it and benchmarks to compare against GpuUInt128. GpuRational should work in a similar way to ERational to avoid excessive memory usage. You have many methods available in GpuUInt128 and GpuUInt128Extensions. Many are also similar to UInt128Extensions and/or ULongExtensions.
+- Prototype a UInt128-native Pow2MontgomeryModWindowed path that matches the GPU implementation once a faster modular multiplication helper is available, so we can revisit removing the GPU struct dependency without regressing benchmarks.
+- Mirror the complete CPU heuristic cycle algorithm on the GPU path so both implementations stay aligned.
 - Extend the GPU numeric type probe to cover proposed custom types such as GpuUInt256 and GpuUInt512 once their definitions are available so we can plan the wide-order heuristics on the device.