trunk/938df06eaa49e465a70514f7fb727bfaedaa6f1f: [DeviceMesh] Make the non-overlapping check more strict (#172343)

The PyTorch team has merged a subtle but significant technical pull request (#172343) that makes the non-overlapping check for DeviceMesh configurations "more strict." This change addresses two critical issues: first, it dramatically speeds up validation from O(world_size) complexity to O(layout length), potentially making checks 100x faster for large clusters. Second, it eliminates support for confusing, non-intuitive DeviceMesh layouts that couldn't be constructed through standard tensor operations like view(), permute(), and flatten().

The fix specifically targets the performance bottleneck that was preventing PyTorch from fully leveraging NVIDIA's CuTe (Collective Communication Thread Execution) layouts, a key technology for optimizing distributed training across thousands of GPUs. By ensuring DeviceMesh follows logical construction patterns—starting with arange(world_size) and applying standard transformations—developers get more predictable behavior when scaling AI models. The PR disallows problematic layouts like (2,3):(3,2) that created irregular rank mappings, which could lead to subtle bugs in large-scale distributed training scenarios.

This technical refinement represents the kind of foundational work that enables the next generation of AI model training. As models grow to trillions of parameters requiring thousands of GPUs, every optimization in the distributed computing stack matters. The PyTorch team's focus on these low-level details ensures the framework remains competitive for enterprise-scale AI deployments where training efficiency directly translates to cost savings and faster innovation cycles.