PyTorch Dynamo fixes descriptor order bug in UDOV setattr tracking

PyTorch's Dynamo JIT compiler has received a targeted fix for how it handles attribute mutations on user-defined objects (UDOVs). When Dynamo traces a Python program, it records operations like setattr, getattr, and delattr to create a computation graph. However, it previously did not respect Python's descriptor protocol ordering when replaying these mutations. In CPython, writes to an object's attribute that go through a descriptor (like a property) follow a specific precedence rule: data descriptors (e.g., properties with both getter and setter) take priority over instance __dict__ entries, while non-data descriptors (e.g., functions) are shadowed by instance attributes. Dynamo's residual replay (the fallback execution path when tracing cannot fully capture behavior) incorrectly treated all mutations as direct __dict__ writes, bypassing descriptor semantics and leading to subtle bugs.

The PR (authored by Codex) introduces side-effect tracking that distinguishes between descriptor-aware mutations and direct instance dict writes. Direct __dict__ writes now replay through descriptor-bypassing helpers (simulating obj.__dict__['attr'] = val), while descriptor-backed and generic mutations keep the normal STORE_ATTR/DELETE_ATTR paths. This ensures that residual replay behaves identically to CPython in all cases. The update also adds focused test coverage for readonly properties, property deleters, non-data descriptor lookup after a `delattr`, __dict__ replacement, and descriptor versus instance-dict precedence. For PyTorch developers relying on Dynamo to compile models using custom Python classes with descriptors, this fix removes a class of hard-to-debug inconsistencies, improving both compilation accuracy and model performance predictability.