fix: remap bare Sequential keys to net.* when loading ddg_da.pt

fix: patch torchmetrics._apply using CPU probe tensor instead of self.device

fix: use callable map_location to skip Lightning model.to() — avoids torchmetrics CUDA assertion

fix: patch raw checkpoint to replace all CUDA refs before load

fix: patch torchmetrics._apply for CPU-only load of GPU-trained checkpoint

fix: track app/models/ package, ignore only top-level /models/