Patch release for the v0.4.0 PyTorch FX graph-to-RTL upgrade.\n\n- Normalizes the PyTorch wheel version so +cpu and +cu build suffixes do not alter canonical manifests\n- Preserves 20/20 compiled graphs, 624 checked RTL words, and 10/10 graph/compiler coverage points\n- Public full release-evidence workflow passes, including report reproducibility\n- Functional accelerator behavior and measured results are unchanged from v0.4.0