Verify PyTorch native operator and symbol workflows

[itdevwu]

Goal

Verify that xprobe can form useful bounded measurements around real PyTorch CPU and CUDA work instead of only synthetic launch fixtures.

Acceptance criteria

• Tests cover representative CPU native ops, CUDA matmul or convolution, memory transfer, and stream synchronization.
• Eager and torch.compile or Triton paths document which native symbols, CUDA APIs, and kernels are stable enough to select.
• Resolver output handles mapped CPython, native-extension, and libtorch objects and presents C++ symbols in agent-usable form.
• Measurements report selector quality, unmatched events, collection pressure, and slowdown against an uninstrumented baseline.
• PyTorch Profiler comparison runs separately when CUPTI subscriber ownership would conflict.
• Native symbol support is not described as Python frame or module.qualname event support.