Make file-based SMC storage MPI-safe by guarding save_step on rank 0

[Copilot]

When running with mpirun and a non-parallel kernel (e.g., VectorMCMC), @rank_zero_run_only on SamplerBase._save_step catches an AttributeError (no _rank on the kernel) and falls back to calling storage.save_step() on all ranks simultaneously, causing concurrent HDF5/Pickle writes and crashes. The workaround was an awkward rank-conditional context manager split.

Changes

• smcpy/utils/mpi_utils.py — adds is_mpi_rank_zero(): checks mpi4py.MPI.COMM_WORLD.Get_rank() == 0, returns True if mpi4py is unavailable. Decouples the rank check from the MCMC kernel's _rank attribute.
• smcpy/utils/storage.py — HDF5Storage.save_step and PickleStorage.save_step early-return when is_mpi_rank_zero() is False. The storage is now independently MPI-safe regardless of kernel type.
• Tests — coverage for is_mpi_rank_zero() across no-MPI, rank-0, and rank-N cases; parametrized tests confirming save_step is a no-op on non-zero ranks for both file storage backends.

With this change, the desired single-context API works with MPI:

# Before: required rank-conditional context split
with HDF5Storage(H5_FILE, "a"):          # load state on all cores
    smc = AdaptiveSampler(kernel)
with HDF5Storage(OUT_FILE, "a") if comm.Get_rank() == 0 else nullcontext():
    steps, mll = smc.sample(N_PARTICLES, N_MCMC_SAMPLES, target_ess=0.9)

# After: single context works across all ranks
with HDF5Storage(H5_FILE, "a"):
    smc = AdaptiveSampler(kernel)
    steps, mll = smc.sample(N_PARTICLES, N_MCMC_SAMPLES, target_ess=0.9)