PyTorch splits test_masked for OOT backends with TestMaskedGeneric

PyTorch just merged a testing infrastructure update from contributor KarhouTam that makes masked-operation tests work across diverse hardware backends. PR #191290 isolates test_where into a new test class called TestMaskedGeneric, explicitly designed to not require a specific device. This means out-of-tree (OOT) backend developers—those building custom accelerators or vendor-specific PyTorch integrations—can now run these tests natively without patching device assumptions.

The refactor also introduces HardwareClassification labels to better categorize tests by target hardware, making it easier to identify which backends are covered. Approved by reviewers fffrog, can-gaa-hou, and albanD, this is a behind-the-scenes quality-of-life improvement for the PyTorch ecosystem. While it doesn't change model performance, it reduces friction for hardware vendors and CI pipelines, ensuring masked operations behave consistently regardless of the underlying chip.