Introduce standalone SkalaXC with ABI-isolated CPU, CUDA, MPI, and language bindings

[Jens (JensWehner)]

Summary

This PR introduces SkalaXC, a standalone C++17 shared library for evaluating Skala exchange-correlation functionals. It reuses GauXC’s lower-level numerical infrastructure while keeping GauXC, LibTorch, Eigen, and other implementation dependencies behind a stable, ABI-isolated public boundary.

The change also replaces the previous GauXC integration examples and Python export tooling with maintained C++, C, Fortran, and Python SkalaXC interfaces, dedicated packaging, and comprehensive CI coverage.

Standalone SkalaXC library

• Adds host-side UKS evaluation for:
  • XC energy
  • scalar and spin-density XC potentials
  • XC nuclear gradients
• Loads and owns the selected TorchScript model for the lifetime of each integrator.
• Uses GauXC’s grid, load-balancing, collocation, molecular-weight, and local-work infrastructure.
• Keeps model features, derivative propagation, matrix assembly, and Skala-specific transformations within SkalaXC.

Public APIs and ABI isolation

• Adds C++, C, and Fortran interfaces.
• Adds a nanobind-based Python binding.
• Keeps public matrices column-major and uses caller-owned output buffers at the C and Fortran boundaries.
• Embeds private dependencies and restricts exported symbols through hidden visibility, linker export maps, and archive exclusion.

Host, MPI, and CUDA execution

• Adds OpenMP-parallel host evaluation.
• Assigns complete atomic grid domains to MPI ranks and performs only final result reductions.
• Broadcasts model archive bytes from communicator rank zero so only that rank needs filesystem access to the model.
• Adds optional CUDA kernels and device-side feature, potential, and gradient orchestration.

CUDA and host-compiler compatibility

• Separates the CUDA build toolchains from unrestricted CPU toolchains.
• Locks the currently validated combinations:
  • CUDA 12.8/12.9 with GCC 14
  • CUDA 13.x with GCC 15
  • CUDA 13.x with Clang 20
• Constrains cuda-compiler to the same major family as the CUDA headers and runtime libraries, preventing mixed CUDA 12 headers with CUDA 13 nvcc.- Applies a temporary Clang compatibility patch to a generated build-tree copy of GauXC’s grid_to_center.cu, removing the C++17-invalid register specifier without modifying the pinned GauXC submodule.

Packaging and dependency management

• Adds Pixi environments for host, MPI, CUDA 12, CUDA 13, Clang, packaging, and development tools.
• Adds a Rattler Build split-package recipe producing:
  • libskalaxc
  • skalaxc-fortran
  • skalaxc-python
• Keeps CUDA 12 and CUDA 13 in separate custom Pixi platform solutions.
• Bundles the validated baseline and Skala model archives.
• Adds relocatable CMake package metadata and separate core/full installation components.

Testing and CI

• Adds Catch2 unit and integration coverage for:
  • model loading and reuse
  • host and CUDA evaluation
  • LDA, PBE, and TPSS behavior
  • matrix views and reordering
  • spin transformations
  • gradients
  • diagnostics
  • OpenMP stress behavior
  • MPI subcommunicators
• Adds black-box C++, C, and Fortran API tests.
• Adds CPU and GPU parity tests against the existing PySCF/GPU4PySCF implementations.
• Adds ABI checks for exported symbols and public consumer dependencies.
• Adds sanitizer, Clang release, clang-tidy, Doxygen, host, MPI, CUDA, and package workflows.

Documentation and examples

• Adds comprehensive architecture, build, API, diagnostics, threading, MPI, CUDA, and packaging documentation.
• Adds minimal public API examples for C++, C, and Fortran.
• Updates the LibTorch and FTorch examples to use the revised model and interface contracts.
• Updates repository-level installation and model-card references.

Legacy cleanup

• Removes the old gauxc Python export package, duplicated GauXC documentation, and external integration examples.
• Replaces the old GauXC example workflow with dedicated SkalaXC workflows.
• Keeps GauXC as a pinned source-tree implementation dependency under SkalaXC rather than exposing or modifying GauXC’s public API.

[Jens (JensWehner)]

rename file

[Jens (JensWehner)]

revert this deletion

[Jens (JensWehner)]

clang-format only.

[Jens (JensWehner)]

format only

[Jens (JensWehner)]

format only

[Jens (JensWehner)]

format only.

[Jens (JensWehner)]

rename to .cpp

[Jens (JensWehner)]

nanobind has an eigen header that should be a lot easier.