Load the GPU BiCGSTAB solver at runtime so GPU-enabled wheels import without CUDA

[FilipovicLado]

Problem

Building with VIENNALS_USE_GPU=ON linked libcudart/libcusparse into
_core.so as DT_NEEDED entries. The dynamic loader resolves those when Python
dlopens the extension module so on a machine
without a CUDA runtime import viennals failed outright:

ImportError: libcudart.so.12: cannot open shared object file

This is what broke the 5.8.3 PyPI wheel.

Approach

Move the CUDA kernels into a separate ViennaLS_GPU shared library and open it
lazily with dlopen() on first use, through a plain C ABI
(lsOxidationBiCGSTABAbi.hpp).

• _core.so no longer references CUDA at all, so it loads anywhere.
• If the library or the CUDA runtime is missing, allocGpuBuffers() returns
  null and the existing gpuIsValid() guards fall through to the CPU solver.
• The nine entry points keep their names, so no call site changed.

New API: GpuMode::Auto

GpuMode::Gpu is documented as "fail if unavailable" and does exactly that —
it raises an error and aborts. GpuMode::Auto is added for the common case:
use the GPU when it's usable, otherwise warn and fall back to the CPU.
GpuMode::Gpu keeps its existing contract. Default remains GpuMode::Cpu.

GPU setup failure paths were restructured into else-if chains. They previously
relied on VIENNACORE_LOG_ERROR throwing to stop execution; with a
non-throwing warning path they would have called into the GPU with a null
handle.

Notes

• Wheels remain GPU-capable, not self-contained: CUDA is still not bundled,
  so the GPU engages only for users who already have a CUDA toolkit installed.
  Everyone else gets the CPU solver instead of a failed import.
• Version bumped to 5.8.5.