Supply CUDA math libs via CUDAToolkit imported targets

[joshkamm]

Fixes #98. Closes #93.

Problem

pixi-build-cmake 0.3.12+ absorbed rattler-build 0.62.0's strict build-environment isolation: the build no longer inherits the developer's shell env (PATH stays on the forwarding allowlist; CPATH does not). nvc++ is still found, but the NVHPC module's math_libs/include no longer reaches it, so every USE_ACC translation unit fails on #include <cublas_v2.h>. The group was shielded by the <=0.3.8 backend cap until #97 bumped it to <=0.4.2 (2026-07-01) — which is why builds "the same as before" suddenly broke. Full evidence trail in #98 (comment).

Fix

Make the build supply the CUDA math libraries itself instead of relying on ambient CPATH:

• find_package(CUDAToolkit REQUIRED) under USE_ACC (top-level)
• target_link_libraries(SlaterGPU PUBLIC CUDA::cublas CUDA::cusolver) — PUBLIC because include/cuda_util.h includes the cuBLAS/cuSolver headers when USE_ACC, so dependents (io, sgpu.exe, downstream consumers like ZEST) need the headers and libraries too; they now inherit both transitively
• examples/: drop the removed-in-CMake-4 FindCUDA module, enable_language(CUDA), and the hand-rolled ${CUDA_TOOLKIT_ROOT_DIR}/../../math_libs/lib64/libcublas.so paths (FindCUDAToolkit understands the NVHPC layout natively); keep CUDA::cudart
• SlaterGPUConfig.cmake.in: replace the deprecated find_package(CUDA QUIET) / enable_language(CUDA) block (the CMP0146 warning from Replace deprecated FindCUDA module with find_package(CUDAToolkit) in Config.cmake.in #93) with find_dependency(CUDAToolkit REQUIRED) gated on the exported USE_ACC setting

Verification

Cold pixi build (fresh cache, wiped .pixi) pinned to backend ==0.3.14 (sanitized env) on Perlmutter, NVHPC 25.5:

• CUDAToolkit 12.9.41 discovered from the NVHPC install via PATH alone (no CPATH)
• All ~20 previously failing cuBLAS/cuSolver TUs compile (cublas_v2.h error count: 0)
• The build then stops at the pre-existing, unrelated NVHPC-25.5 NVVM failure in integrals.cpp (OpenACC kernel names truncated for long source paths, causing duplicate device functions (NVVM_ERROR_COMPILATION) #99), which this login node hits regardless of this change — so the final sgpu.exe link is not exercised locally. CI's GPU runner (NVHPC 25.1) should exercise the full link path on this PR.

Notes for review

• The CMake ≥4.1 warning at the top of CMakeLists.txt (from Fix CI after broken from new cmake release #50) referenced CUDART_LIBRARY errors produced by the old FindCUDA/find_library code in examples/, which this PR removes — the warning may now be obsolete, but I left it for a separate decision (the monthly latest-pixi CI run can re-evaluate it).
• Consumers (ZEST, XCtera) need no changes: the exported SlaterGPUTargets.cmake now carries CUDA::cublas/CUDA::cusolver in the interface, resolved by the Config file's find_dependency(CUDAToolkit). Their lock hashes will need the usual bump after merge.

🤖 Generated with Claude Code