Fixes the build failure when configuring with POLYSOLVE_LARGE_INDEX=ON

[Chenyang-Joe]

With large indices, StiffnessMatrix uses 64-bit (std::ptrdiff_t) indices, but two of the wrapped direct solvers are 32-bit int-only and fail to compile against it:

• Apple Accelerate — the Accelerate sparse API is int-only, with no 64-bit variant. Disabled in CMake when POLYSOLVE_LARGE_INDEX is on, so the macro is never defined and no Accelerate code is compiled. (This is the macOS compile error seen on the large-index build.)
• MKL Pardiso — Eigen's PardisoSupport only specializes pardiso_run_selector for int/long long, not long (std::ptrdiff_t). Excluded under #ifndef POLYSOLVE_LARGE_INDEX, mirroring the existing UmfPack handling. MKL's BLAS/LAPACK backend is left enabled.

This PR only makes the large-index build compile — it does not add large-index support to those solvers (that's future work). SuperLU has the same int-only limitation and is left untouched here.

Testing

• Linux: library builds and the unit tests pass with POLYSOLVE_LARGE_INDEX=ON.
• Apple Silicon (Accelerate on by default): the large-index build now compiles — both polysolve standalone, and inside a polyfem build that points its polysolve dependency at this branch.

@maxpaik16 PTAL! Thanks