Test the qiskit-addon-sqd entry points, serially and under MPI

[garrison]

Nothing in the suite covered the solver wrapper that qiskit-addon-sqd is
meant to be handed. test_reference_energies goes through
tpb_diag_from_files, and the only place diagonalize_fermionic_hamiltonian
ran was run_sqd_sbd.ipynb under nbmake — which asserts nothing, so it
caught a raised exception and nothing else, on a single rank.

What this adds

test/test_sqd_integration.py, three checks each in a _standalone and an
_mpi variant sharing one body:

Test	Subspace	Assertion
test_small_subspace_*	fixed, 40 dets	energy pinned to -85.29400074571684
test_published_energy_*	full 1em3 selection	the published -76.23594663 (slow)
test_diagonalize_fermionic_hamiltonian_*	sampled by upstream	bracketed, plus shape/occupancy consistency

The first two are deterministic: the determinants are given rather than
sampled, so the answer does not depend on the process count, which is the
property worth pinning. The third runs the real self-consistent loop, so a
change to the sci_solver contract or to SCIResult/SCIState fails here
instead of in a user's script; its energy is bracketed because the subspace
comes from upstream's sampling and recovery.

All three leave fcidump_path unset, so rank 0 regenerates the FCIDUMP into
a temp dir and broadcasts the path for every rank to open — the default, what
a caller coming through diagonalize_fermionic_hamiltonian gets, and the
multi-rank path 7c7cc09 fixed.

Verified identical to ten digits on 1, 2 and 4 ranks.

Why two variants instead of one test

pytest-mpi filters on the mpi marker in opposite directions: --only-mpi
skips what is not marked, and no flag skips what is. So one function cannot
run in both modes. The bodies size the alpha-determinant grid from
MPI.COMM_WORLD, which is 1 in a single process, so the same code serves
both — tox -e py runs the _standalone variants and tox -e mpi the
_mpi ones, with no new tox env or CI plumbing needed.

bit_length

Set to 63 rather than the wrapper's default of 20. SBD fixes the packed word
count in a process-wide inline static on the first diagonalization and throws
det_vector: elem_size mismatch for any later one implying a different
count, so every test sharing a process must agree; 63 gives h2o the single
word that test_reference_energies already gets from 64.

63 is also the largest legal value. As #12 established, bitadvance()
computes (((size_t) 1) << bit_length) - 1, so 64 shifts a 64-bit size_t
by 64 — undefined behavior, reached from mpi_redistribution() and
mpi_sort_bitarray(), which is exactly what the _mpi tests exercise.

Two known gaps, both noted in the module:

• Multi-word packing goes uncovered, and that is what a caller taking
  SBD_DEFAULT_BIT_LENGTH (20) actually gets. Covering it needs a module
  that does not share a process with these.
• test_reference_energies still pins 64. Same UB, but on
  tpb_diag_from_files rather than the MPI paths, and out of scope here.

Not a performance question either way: at 275 determinants the solve takes
1.9s at 63 against 2.1s at 20.

Other changes

• conftest.py: device_config and counts_path fixtures. The solver
  wrappers take a DeviceConfig rather than a backend module, so the
  "was this backend built?" skip is factored into a helper both fixtures
  call instead of being duplicated.
• pyproject.toml: qiskit-addon-sqd and pyscf move into the test
  extra so tox -e py and tox -e mpi both cover this path. The tests
  skip themselves if the imports are unavailable.

---

This PR was generated by Claude Opus 5 under my guidance.