Native AMICA build + cross-platform release workflow (#165, phases 1-2)

[neuromechanist]

Summary

First two phases of epic #165: a dependency-free native AMICA build (single-rank MPI shim) and the cross-platform release workflow. Purely additive: native/ build tooling and a release-triggered workflow, with no change to the Python package or its CI. Merged early (per plan) so the 5-platform release matrix can be dispatch-validated before the Python native engine (Phase 3) is built on top.

Phases 3 (native run engine + binary resolver) and 4 (rename to pamica) remain open under epic #165.

Phase 1 (#166, merged to epic in PR #170)

Single-rank MPI shim (native/mpi_single.{f90,c}) + portable random_seed + vmath_shim remove the Open MPI / MKL / AMD-LibM runtime dependencies, so the AMICA Fortran reference builds into a self-contained binary. Proven mathematically identical to real Open MPI at machine epsilon (iteration-1 LL agrees to 8.9e-16; native/validate_shim.sh).

Phase 2 (#167, merged to epic in PR #171)

.github/workflows/release-binaries.yml builds the recipe natively on each of macos-arm64, linux-x64, linux-arm64, windows-x64 and windows-arm64, smoke-tests each binary on real sample EEG, and attaches checksummed assets to the release. The macOS binary is verified fully portable locally (links only Accelerate + libSystem).

Test plan

• native/validate_shim.sh: PASS at 8.9e-16 (shim == Open MPI).
• macOS arm64 build: otool -L shows only system libraries.
• The Linux/Windows/ARM matrix validates on the first workflow_dispatch after this reaches main (that is the reason for the early merge). windows-arm64 is experimental (continue-on-error).