Phase 1: dependency-free native build (single-rank MPI shim)

[neuromechanist]

Summary

Builds the AMICA Fortran reference into a self-contained binary with no MPI runtime, no MKL, and no vendor math library, so it can ship as a release asset and run without a user-installed toolchain. Extends sccn/amica PR #53's vendor-neutral recipe with a single-rank MPI shim.

Closes #166. Part of epic #165.

Key result

amica15.f90 is MPI+OpenMP+LAPACK but pyAMICA runs it as one process; every MPI collective it uses is trivial for a single rank. native/mpi_single.{f90,c} provide the constants + ~11 subroutines (reduce/gather become memcpy(recv, send, count*datatype)), removing the Open MPI runtime dependency (the last one after PR #53 removed MKL/AMD LibM).

Proven mathematically identical to real Open MPI: validate_shim.sh builds the same patched source with the shim (gfortran) and with mpif90, pins the RNG seed, runs one iteration on the bundled sample EEG, and asserts machine-epsilon agreement: LL max|diff| = 8.9e-16, W/A = 1.7e-18, all params <= 5e-16 (tol 1e-12). One iteration is the discriminating window (the optimizer amplifies last-bit differences over iterations, cf. #51/#27). The shim binary links no libmpi.

Test plan

• bash native/validate_shim.sh (needs gfortran + mpif90): PASS at 8.9e-16.
• Static LAPACK linking and the 5-platform build matrix are Phase 2 (Phase 2: cross-platform release CI (5 targets attached to release) #167).