Summary

Fixes winml perf --module <ClassName> on HuggingFace models like microsoft/resnet-50, which previously failed with AttributeError: <Parent> has no attribute '<...>' and, once that was past, forward() missing 1 required positional argument.

Two root causes:

1. Wrong parent class instantiated. _perf_modules passed the CLI task (typically None) to _instantiate_parent_model. When the model_type's first supported task differs from the architectures-derived task used by generate_hf_build_config, the parent class mismatches the recorded module_path. For microsoft/resnet-50, build configs use ResNetForImageClassification (paths prefixed with resnet.), but perf was instantiating bare ResNetModel → AttributeError: ResNetModel has no attribute 'resnet'. Fix: re-resolve the parent task from the model_id via resolve_loader_config, matching the path used to compute module_path.

2. Wrong input tensor names. _build_submodule_config hardcoded input names as input_0, input_1, …, but PyTorch submodules expect specific positional/keyword names (e.g., ResNetStage.forward(input), ResNetBottleNeckLayer.forward(hidden_state), ResNetEmbeddings.forward(pixel_values)). Calling submodule(input_0=tensor) raises forward() missing 1 required positional argument. Fix: propagate forward-arg names through SubmoduleInfo.input_names (sourced from the existing hook-capture data, with a inspect.signature fallback) and use them when building InputTensorSpec. Falls back to input_{i} only when names can't be discovered.

End-to-end on a CPU box:

$ winml perf -m microsoft/resnet-50 --module ResNetStage --ep cpu --device cpu
                Per-Module Perf: ResNetStage
| Module Path             | Mean (ms) | P90 (ms) | Min  | Max  |
| resnet.encoder.stages.0 |      4.16 |     5.22 | 3.30 | 5.22 |
| resnet.encoder.stages.1 |     14.04 |    24.09 | 8.69 | 24.09 |
| resnet.encoder.stages.2 |      7.41 |     8.65 | 4.93 | 8.65 |
| resnet.encoder.stages.3 |      5.20 |     7.09 | 3.59 | 7.09 |

Closes #192