examples: document the JAX/MPI conflict in the enlarge-subspace driver, unify GPU examples

[Sophia Wen (hfwen0502)]

Summary

Documents a real failure mode in run_sqd_enlarge_subspace_sbd.py, and makes
the GPU example commands consistent.

• JAX/MPI conflict. The excitation-expansion step between rounds goes
  through qiskit-addon-sqd's own JAX-based
  enlarge_batch_from_transitions, which has no MPI awareness. Every rank
  runs it redundantly, and if JAX is set up for GPU, every rank tries to
  grab a GPU for it at once and the run dies. Documented in
  python/examples/README.md and in the driver's own module docstring, with
  the fix: run on a single rank, or set JAX_PLATFORMS=cpu (SBD's own
  --device gpu diagonalization is unaffected either way).
• Consistent GPU examples. All --device gpu example commands now use
  8 MPI ranks with --adet_comm_size 4 --bdet_comm_size 2, and none set
  --task_comm_size.
• AMD note. The Overview now spells out which --device value each
  backend maps to, and says to substitute --device gpu-omp for
  --device gpu on AMD GPUs.

Verification (on h100, 8x H100)

The JAX/MPI premise was confirmed directly rather than assumed — h100's JAX
is now GPU-enabled (jax.devices() reports 8 CudaDevices):

• Without JAX_PLATFORMS=cpu, 8 ranks, --device gpu: fails with
  jax.errors.JaxRuntimeError: RESOURCE_EXHAUSTED: CUDA_ERROR_OUT_OF_MEMORY.
• With JAX_PLATFORMS=cpu, same command: converges to
  -76.2421767512, matching the documented value.
• run_sqd_sbd.py's H2O example at the new 8-rank / 4x2 / --device gpu
  settings: -76.2359466308, matching its documented value.
• The driver docstring's command, run verbatim: -76.2421767512.
• JAX_PLATFORMS=cpu mpirun ... (prefix form, as documented) does propagate
  to all ranks under this MPICH build, so no -x/-genv is needed.
• gpu-omp confirmed a valid --device choice in all three drivers'
  argparse, and the AMD guidance matches device_config.py's own docstrings
  (gpu() is "NVIDIA only ... on an AMD host use gpu_omp instead";
  gpu_omp() "Works on NVIDIA and AMD").

Not tested on AMD hardware — the AMD note restates this repo's own
documented backend support rather than a fresh measurement.

Test plan

• Both documented energies reproduced at the new rank/grid settings.
• Failure mode reproduced, and the documented workaround shown to fix it.
• Driver docstring command runs verbatim.
• Markdown fences balanced; no example command sets --task_comm_size.