Weak references in dump_allocated_tensors may crash profiler during training

[perctrix]

gc.get_objects() in dump_allocated_tensors() can return weakref proxy objects. A proxy passes isinstance(obj, torch.Tensor) (forwarded __class__), but if the referent is collected before .numel() / .element_size() / .shape are accessed, a ReferenceError is raised.

This is called from Profiler.record_allocated_tensors() during training (training.py:502, 544), so a crash here interrupts a running experiment.

Fix: wrap tensor attribute access in try/except ReferenceError.