v0.3.0 — prebuilt CUDA cores, nothing compiles after pip install

sweep-solver 0.3.0

pip install -U sweepx and the CUDA backend works — nothing compiles.

• Prebuilt CUDA cores in the wheel. One torch-free core per CUDA major:
  cu12 (V100 through H100/H200, Blackwell via PTX) and cu13 (T4/RTX 20 and
  newer, Blackwell native, driver >= 580). impl='c' drives it through a
  pure-Python ctypes layer, so no nvcc, C++ compiler or CUDA headers are needed
  after pip install, and the same wheel works with any torch version and
  Python >= 3.9. An nvcc is needed only when no shipped core fits (then only the
  core is built, once).
• The compiled CPU engine is gone. CPU runs use the eager backend.
• VRZ gradients are exact: the CUDA adjoint is the discrete transpose of the
  CPML forward; every memory mode images U_it through p_tt; a sigma=0
  boundary_buffer keeps the imaging stencil inside reconstructed cells; a
  persistent boundary-saving runner for AcousticVRZ3D.
• Boundary saving is lean again for VRZ, first-order VTI, ElasticTTI2nd and
  DASZhao: the backward no longer holds an nt-long replay buffer.
• Correctness fixes: qP VTI/TTI model gradients were all NaN; 2-D
  AcousticVTI1st swapped dx/dz on anisotropic grids; propagators sharing an
  equation could step with each other's CPML profiles; DD gather_record
  crossed shot groups and DD accepted multi-shot geometry it cannot answer;
  2-D VRZ boundary saving started its reverse pass from a damaged state.
• Domain decomposition: one batched P2P per halo shipment, NCCL timeout
  honoured by sub-groups.

Full list: CHANGELOG.

Companions: sweep-agent 0.0.5 (admits solver 0.3), sweepx 0.3.0
(sweep-solver>=0.3,<0.4, sweep-agent>=0.0.5).

PyPI: sweep-solver 0.3.0 · sweep-agent 0.0.5 · sweepx 0.3.0