Validate MPI reduction of Enzyme gradients

[krystophny]

Depends on #47.
Refs #15.

Summary

• add an MPI-C shim for Enzyme validation so the test does not depend on compiler-specific Fortran MPI modules
• add a two-rank Enzyme gradient test where each rank differentiates a local energy contribution and MPI reduces the scalar energy and full gradient outside __enzyme_autodiff
• register the MPI validation only when MPI C and mpiexec are available

Scope

This validates the AD contract needed for MPI: Enzyme sees only rank-local energy work, and MPI collectives happen outside the differentiated region.

It does not claim full USE_MPI=ON support for the installed LLVM flang stack. On this machine, flang-new cannot configure a Fortran MPI build because the available MPI Fortran modules are not LLVM-flang modules.

Verification

Test fails on main

Before this PR, the MPI Enzyme reduction target is absent on the stacked parent:

$ cmake --build <parent-build> --target gvec_enzyme_energy_gradient_mpi
build target exit=1
ninja: error: unknown target 'gvec_enzyme_energy_gradient_mpi'

The full USE_MPI=ON flang route is also blocked by the local MPI Fortran toolchain, and this PR does not hide that:

$ cmake -S . -B <build-mpi-fortran> -G Ninja -DCMAKE_Fortran_COMPILER=flang-new -DCMAKE_BUILD_TYPE=Release -DGVEC_FIX_HMAP=RZ -DUSE_OPENMP=OFF -DUSE_MPI=ON -DLINK_GVEC_TO_NETCDF=OFF -DUSE_ENZYME=ON
cmake USE_MPI=ON exit=1
-- The Fortran compiler identification is LLVMFlang 22.1.6
-- Found MPI_C: /usr/lib/libmpi.so (found version "3.1")
-- Found MPI_CXX: /usr/lib/libmpi.so (found version "3.1")
-- Could NOT find MPI_Fortran (missing: MPI_Fortran_WORKS)
Could NOT find MPI (missing: MPI_Fortran_FOUND) (found version "3.1")
-- Configuring incomplete, errors occurred!

Test passes after fix

$ cmake -S . -B <build> -G Ninja -DCMAKE_Fortran_COMPILER=flang-new -DCMAKE_BUILD_TYPE=Release -DGVEC_FIX_HMAP=RZ -DUSE_OPENMP=OFF -DLINK_GVEC_TO_NETCDF=OFF -DUSE_ENZYME=ON
-- The Fortran compiler identification is LLVMFlang 22.1.6
-- Configuring done
-- Build files have been written to: <build>

$ cmake --build <build> --target gvec_enzyme_energy_gradient_mpi -j$(nproc)
[1/4] Building C object CMakeFiles/gvec_enzyme_mpi_shim.dir/test-CI/enzyme/gvec_enzyme_mpi_shim.c.o
[2/4] Linking C static library lib/libgvec_enzyme_mpi_shim.a
[3/4] Generating gvec_enzyme_energy_gradient_mpi.enzyme/gvec_enzyme_energy_gradient_mpi.o
[4/4] Linking Fortran executable bin/gvec_enzyme_energy_gradient_mpi

$ ctest --test-dir <build> -R '^gvec_enzyme_energy_gradient_mpi$' --output-on-failure
1/1 Test #4: gvec_enzyme_energy_gradient_mpi ...   Passed
100% tests passed, 0 tests failed out of 1

$ ctest --test-dir <build-all> -R 'gvec_enzyme_(smoke|array_gradient|rz_flat_gradient|energy_gradient_mpi|gvec_module_smoke|ad_fixture_gradient)$' --output-on-failure
1/6 Test #1: gvec_enzyme_smoke .................   Passed
2/6 Test #2: gvec_enzyme_array_gradient ........   Passed
3/6 Test #3: gvec_enzyme_rz_flat_gradient ......   Passed
4/6 Test #4: gvec_enzyme_energy_gradient_mpi ...   Passed
5/6 Test #5: gvec_enzyme_gvec_module_smoke .....   Passed
6/6 Test #6: gvec_enzyme_ad_fixture_gradient ...   Passed
100% tests passed, 0 tests failed out of 6

$ git diff --check
<no output>

$ fo
fo: no Fortran project detected

[krystophny]

Added the AD-14 MPI reduction validation slice.

The fix adds a two-rank MPI-C validation target where Enzyme differentiates only a rank-local flat-kernel energy contribution. The scalar energy and full gradient are reduced with MPI outside __enzyme_autodiff, and the reduced result is compared against a serial full-domain Enzyme gradient.

The PR intentionally avoids Fortran MPI modules in the Enzyme path because the local LLVM flang toolchain cannot consume the installed Fortran MPI modules. Full USE_MPI=ON Enzyme support still needs an LLVM-compatible Fortran MPI stack or a larger MPI abstraction change.