Add UMFPACK sparse solver as PARDISO alternative for non-Intel platforms

[mahopan]

Motivation

WinNet requires Intel MKL PARDISO for sparse linear solves in netsolve(), which limits portability to Intel-compiler platforms. This PR adds SuiteSparse/UMFPACK as a drop-in alternative, enabling WinNet to build and run on Apple Silicon Macs (and other non-Intel platforms) using gfortran + Homebrew.

Changes

No existing WinNet source files are modified. Three new files are added:

src/external_tools/umfpack_interface.f90

Fortran 2003 iso_c_binding wrapper for the UMFPACK C API. Wraps umfpack_di_symbolic, umfpack_di_numeric, umfpack_di_solve, umfpack_di_defaults, and the memory-free routines.

src/external_tools/pardiso_shim.f90

Drop-in replacement for Intel MKL's pardiso() subroutine, matching the exact API signature that netsolve() calls. Provides two compile-time backends:

Backend	Flag	Dependencies
UMFPACK (default)	-DUSE_UMFPACK	SuiteSparse (-lumfpack -lamd -lsuitesparseconfig)
Dense LAPACK	-DUSE_DENSE_SOLVER	BLAS/LAPACK only (no external sparse solver)

Key technical detail: CSC/CSR transpose

WinNet's netsolve() constructs the Jacobian in CSC format (ia = column pointers, ja = row indices) but passes it to PARDISO using PARDISO's CSR calling convention (ia = row pointers, ja = column indices). PARDISO therefore interprets the CSC data as CSR and effectively solves A^T · x = b. This is not a bug — the Newton-Raphson iteration converges correctly with A^T — but any replacement solver must reproduce this transposed solve:

• UMFPACK backend: uses umfpack_di_solve(UMFPACK_At, ...) (transpose solve)
• Dense LAPACK backend: transposes during matrix assembly (A_dense(j, row))

Makefile.example_mac_arm

Build template for Apple Silicon Macs (aarch64):

• gfortran via h5fc (Homebrew HDF5)
• OpenBLAS (replacing Intel MKL BLAS/LAPACK)
• SuiteSparse/UMFPACK (replacing Intel MKL PARDISO)
• Auto-detects sparse solver backend; configurable via SPARSE_SOLVER={UMFPACK|PARDISO|DENSE}

Build Instructions (Apple Silicon Mac)

brew install gcc hdf5 openblas suite-sparse
cp Makefile.example_mac_arm Makefile
make

Test Results

Example	Status	Notes
neutron-decay	✅	30 iterations
NSE (1910 species)	✅	1 iteration
CCSN explosive Si-burning (2919 species)	✅	2874 iterations, 52.6s

The CCSN example reproduces the expected results from Reichert et al. (2023) Fig. 14:

• Iron-group peak at A ≈ 56: X(⁵⁶Fe) = 0.60, X(⁴He) = 0.24, X(⁵⁸Ni) = 0.083
• Σ X = 1.000000 (mass conservation)
• Complete trajectory: T₉ = 6.0 → 5.8 GK (NSE→network) → 0.01 GK (freeze-out)

Scope

This PR is intentionally minimal — it adds a new solver option without touching any existing code. The original Intel/PARDISO build path is completely unaffected.