PyTorch PR #188978 simplifies FakeTensor checks with is_fake_tensor()

PyTorch merged a foundational refactor (PR #188978) that replaces all Python `isinstance(t, FakeTensor)` calls with the new `is_fake_tensor(t)` utility function. The change, authored by liangel-02, introduces `is_fake_tensor` in `fake_tensor.py` as the central detection mechanism. It also adds a lint rule, `ISINSTANCE_FAKE_TENSOR`, to prevent future code from directly using `isinstance` against `FakeTensor`—unless the developer intentionally wants to check for the Python-only subclass. This is currently a no-op: the new function behaves identically to the old `isinstance` check, returning `True` only for Python `FakeTensor` instances.

However, the real purpose is to prepare for a C++ `FakeTensor` implementation that will eventually replace the Python version. The `is_fake_tensor` function is designed to call helper functions that access C++ `FakeTensor` attributes (like `fake_device`, `real_tensor`, `constant`) via pybind, while falling back to `None` for now. Once the C++ type ships, `is_fake_tensor` will seamlessly detect both implementations, enabling a smooth migration. The lint rule ensures that existing `isinstance` patterns are not reintroduced, and the PR notes that this rule should be removed once the C++ rollout is complete. This behind-the-scenes cleanup is critical for PyTorch's ongoing performance optimizations and core infrastructure evolution.