trunk/4b169340ef40f11b42b52925f1329a18a152d9e1: [dynamo] Support retracing inference_mode helpers (#185066)

Generated Dynamo graphs for torch.inference_mode() contain calls to

torch.autograd.grad_mode._enter_inference_mode and _exit_inference_mode.

Retracing those graphs used to treat _enter_inference_mode as an ordinary

in-graph torch function, then failed when handle_traced_output tried to wrap

its non-Tensor _InferenceMode return value.

Teach TorchInGraphFunctionVariable to map those generated helper calls back to

Dynamo's existing InferenceModeVariable. The enter handler now records the same

FX node and cleanup behavior used by the original context-manager path, and the

exit handler delegates to InferenceModeVariable.exit(). This keeps the fix scoped

to the generated helper pair instead of adding a broad non-Tensor escape hatch in

handle_traced_output, which would hide unrelated unsupported returns.

Fixes

Approved by: https://github.com/rtimpe