PyTorch Dynamo gets proper power operator support in PR #186296

Pull request #186296 brings long-overdue support for power operators (`pow` and `ipow`) to PyTorch’s Dynamo compiler. Previously these operators were handled via an older `_handle_op_in_graph` table, which lacked proper integration with Dynamo’s variable tracking system. The PR rewires `operator.pow` and `operator.ipow` through the clean `ternary_op` and `ternary_iop` dispatch in `BuiltinVariable`, ensuring they benefit from Dynamo’s full compile-time optimization pipeline.

To handle Python’s ternary power semantics (which include a modulus argument in CPython’s `nb_power`), the implementation adds `nb_power_impl` (with a `reverse: bool` flag for forward/reflected operations) and `nb_inplace_power_impl` to all major variable trackers: `ConstantVariable`, `TensorVariable`, `SymNodeVariable`, and `UserDefinedVariable`. A separate `nb_power_z_impl` is introduced for the modulus ‘z-slot’ to avoid conflating it with the forward/reverse logic. The PR was co-authored by Claude Sonnet, reflecting growing AI-assisted development in PyTorch core.