examples: fix the enlarge driver's unreachable closure test

[Sophia Wen (hfwen0502)]

Summary

enlarge_via_singles() builds its bitstring rows from the threshold-selected
strings only:

ia_idx, ib_idx = np.nonzero(weights > threshold)
alpha_bits = ints_to_bool_columns(np.asarray(ci_strs_a)[ia_idx], norb)

Row 0 of transitions is the identity, so everything fed in comes back — but
a solved string whose every pair fell below --enlarge_threshold is never fed
in, and never comes back. So the result is a superset of the dominant
strings, not of the subspace. The docstring claimed the opposite ("Always a
superset of the input ci_strs_a/ci_strs_b"), and the loop's closure test
believed it and tested equality.

Equality additionally requires every solved string to be dominant, which is
false on any subspace with a low-weight tail. So no_growth never became
true, the "expansion added nothing new" stop reason was dead code, and every
run paid an extra round to establish energy/occupancy convergence — negligible
on a small molecule, minutes on anything large.

Fix: the subset test the condition always meant, matching
run_sqd_sbd_carryover.py:

no_growth = (set(new_alpha.tolist()) <= set(int(x) for x in ci_strs_a)
             and set(new_beta.tolist()) <= set(int(x) for x in ci_strs_b))

Still computed on the raw expansion, before cap_to_max_dim — capping
first would mask real growth and could stall a run while reporting
convergence. Docstring corrected to describe the pruning. A nearby --max_dim
comment is also made dataset-agnostic.

Verification (h100, 8 GPU ranks, H2O)

main and the fix back to back, same flags:

	rounds	stop reason	energy	final subspace
main	3	energy converged	-76.2421767512	1742 x 1742
this PR	2	expansion added nothing new	-76.2421767512	1742 x 1742

Identical energy to ten digits over an identical final subspace — the fix
changes when closure is recognised, not what the loop converges to.

Round 2 is the bug in miniature: the expansion is 1720 strings against
1742 solved. A strict subset adding nothing — but 1720 != 1742, so
equality declared growth. (Same two numbers #39's description explains as a
quirk of this test; this is the cause.)

Notes for review

• Scoped to this defect. The module docstring still references a
  run_sbd_selected_ci.py that no longer exists; examples: SQD driver that grows its subspace via SBD's own carryover #39 already corrects those,
  so fixing them here would only conflict.
• cap_to_max_dim is untouched. Separately, once the include list reaches
  max_dim it fills upstream's whole budget on its own, since upstream builds
  unique(concat(include_a, include_b, carryover, samples))[:max_dim] — a
  design question, not part of this fix.