trunk/3168fb82bf76c63e1cfed486ba1d0765bb47f559: Defensively deepcopy non-immutable config defaults (#180258)

The PyTorch team has resolved a subtle but potentially dangerous bug in the framework's configuration system that could have silently corrupted AI model training runs. The issue, identified as PR #180258, was in the ConfigModule's handling of mutable default values like lists, sets, and dictionaries. The original implementation used an allowlist approach that only copied these three specific mutable types, creating a fragile system where any new mutable type added as a config default would skip the copy operation entirely. This meant that different instances of a configuration could accidentally share and mutate the same default object, leading to unpredictable behavior in training pipelines.

Authored with assistance from Anthropic's Claude AI, the fix fundamentally changes the defensive copying strategy. Instead of trying to identify which types need copying, the new implementation allows only known immutable scalar types (like integers, strings, and tuples) to pass through without copying. Any unrecognized type is now defensively deep-copied by default, creating a much more robust safety net. This 'defensive by default' approach prevents future developers from accidentally introducing mutable defaults that could be shared across configuration instances, eliminating a whole class of subtle bugs before they can impact production AI systems.

The change represents an important shift in how AI frameworks handle configuration safety, moving from a brittle allowlist approach to a more principled defensive programming pattern. For PyTorch's massive user base of 99.1k GitHub stars, this fix ensures that complex configuration hierarchies in machine learning pipelines remain predictable and isolated, preventing hours of debugging for what would otherwise appear as mysterious training failures or inconsistent model behavior across different runs.