D11 — Own CUDA kernels, no PyTorch or DREAMPlace runtime

Decision

Own CUDA kernels for placement and routing; no PyTorch/DREAMPlace runtime dependency

Alternatives

Embed DREAMPlace/Xplace

Reason

PyTorch wheels dropped Volta; CCCL 3/cuCollections partly untested on Pascal; BSD-3 code is ported, not linked