Helper function for MPI-accelerated GST fits

[rileyjmurray]

This PR adds Protocol.run_mpi() to run MPI-accelerated protocols launched as a subprocess. This function can be called directly from a Jupyter notebook, an interactive Python session, or a plain script. It removes the need for the user to know about mpi4py (besides its existence) and the need to write their own MPI launcher scripts.

Key changes

Added Protocol.run_mpi(...) to pygsti/protocols/protocol.py. Notable features of this function:

• Accepts num_ranks, an optional work_dir, forwarded **run_kwargs, and several keyword arguments for controlling launcher selection, thread counts, and dry-run behavior.
• When num_ranks=1, it short-circuits to a plain self.run() call with no subprocess overhead.
• Auto-detects the MPI launcher (mpiexec, mpirun, or mpiexec.hydra) when mpiexec='auto'.
• Auto-computes BLAS thread counts per rank (OMP_NUM_THREADS, MKL_NUM_THREADS, etc.) based on physical core count, with a blas_threads_per_rank override.
• Supports dry_run=True for generating the launch command and writing data to a permanent work_dir without executing. This is useful for embedding into HPC batch scripts.

This new function is tested in test/integration/test_gst_run_mpi.py and demonstrated in
jupyter_notebooks/Examples/MPI-RunningGST-Automagic.ipynb.

Minor change: tests in test/unit/mpi/test_mpi.py have been made a little more robust. The MPI launcher checks both mpiexec and mpirun. We pass --oversubscribe as appropriate so our use of 4 MPI ranks doesn't break on systems with < 4 physical CPU cores.

Incidental changes

• pygsti/tools/optools.py: extract real part of trace for Hermitian matrices (makes warning messages clearer)
• pygsti/report/colormaps.py, plotly_plot_ex.py: robustness against plotly version differences.
• numpy.core namespace access removed (deprecated in NumPy 2.x).
• Removed a spurious warning about non-MPI-accelerated gauge optimization.