[cuda-unified-memory-analyzer] Add more diagnostic features

In ML pipelines that rely on `cudaMallocManaged`, performance can degrade sharply once allocations exceed what the GPU can keep resident. The tricky part is that the transition from **resident memory → page-fault migration** isn’t visible from typical tooling. I built a small diagnostic tool that identifies that boundary directly. It performs controlled allocation pressure and reports: • GPU **residency limit** • **Fault onset ratio** where migration begins • **Thrash detection** when memory repeatedly migrates Linux [https://github.com/parallelArchitect/cuda-unified-memory-analyzer](https://github.com/parallelArchitect/cuda-unified-memory-analyzer)