Build with gfortran + Open MPI on macOS (incl. Apple Silicon) and Linux, no MKL

[neuromechanist]

Fixes the M1 compile errors reported in #44 and gives a vendor-neutral build path (also relevant to #49's Ubuntu question). amica15.f90 targets Intel's toolchain (ifort + MKL), so a plain gfortran + Open MPI + LAPACK build hits three issues; this PR resolves all three and adds a one-command build script + docs.

Changes

• Portable random_seed (amica15.f90, amica15_header.f90). The source passed a fixed size-2 seed array, which only ifort accepts; gfortran requires an array of its own (larger) size. Now it queries random_seed(SIZE=...) and fills the full-size array from the clock/rank/seed, so per-rank streams still differ. This was the root of several of the "no IMPLICIT type"/compile failures once the earlier errors are cleared.
• vmath_shim.c supplies vrda_exp/vrda_log (the non-MKL branch's vectorized exp/log, otherwise from AMD LibM) as libm loops, matching gfortran's by-reference/trailing-underscore ABI. No vendor math library (MKL VML or AMD LibM) is needed; results are IEEE-accurate.
• build_gfortran.sh compiles with the flags this Intel-oriented source needs under gfortran: -cpp (resolve #ifdef MKL so the mkl_vml.f90 include is skipped), -std=legacy -fallow-argument-mismatch (obsolescent constructs), -ffree-line-length-none, and per-platform LAPACK (Homebrew lapack/Accelerate on macOS, -llapack -lblas on Linux). Runs as a single Open MPI rank (no mpirun); threads via OMP_NUM_THREADS.
• README gains a Compilation section; ships amica15_macos_arm64, a native Apple-Silicon binary (links Homebrew Open MPI + LAPACK, so running it needs brew install gcc open-mpi lapack).

Verified

Builds and runs on macOS arm64 (gfortran 16, Homebrew Open MPI + LAPACK) and Ubuntu 24.04 x86_64 (gfortran 13, apt Open MPI + liblapack-dev). A short fit on real sample EEG converges normally (LL ~ -3.46 at iter 5), matching the Intel build's trajectory.

Toolchain:

• macOS: brew install gcc open-mpi lapack
• Ubuntu/Debian: sudo apt-get install -y gfortran libopenmpi-dev openmpi-bin liblapack-dev libblas-dev

[neuromechanist]

Removed the random_seed source change here (509c8be); that is now owned by #54, which does the same portable random_seed(SIZE=...) fix plus a reproducible seed param. This PR is now the gfortran / no-MKL build tooling only (build_gfortran.sh, vmath_shim.c, README, arm64 binary). Since the seed change is gone from here, building with gfortran needs #54's portable random_seed in the source first, so #54 should merge before this.