AUDIT F11 follow-up: support solvent in geometry_opt with required final single point

[jonathanschultzNU]

Summary

Implements solvent support for geometry optimization calculations by mirroring the interactive app's approximation: optimize in gas phase, then run a required solvated single point on the final geometry. This resolves an inconsistency where the UI enabled the solvent checkbox for geometry_opt but the worker would reject or silently ignore it.

Replaces #121, which targeted claude/gpt-astra-audit-fixes-iqa1om (now merged via #120 and deleted) and can no longer be reopened/retargeted since GitHub won't change the base of a closed PR once its base branch is gone. Same branch (claude/code-review-t1ji0e), now against main.

Key Changes

Worker backend (quantui/backends/worker.py):

• Added "geometry_opt" to _SOLVENT_SUPPORTED_CALC_TYPES (alongside "single_point" and "reorganization_energy")
• Extended _run_geometry_opt() to run a required solvated single-point calculation on the optimized geometry when request.solvent is set
• The solvated single-point energy and convergence status replace the optimizer's final values, ensuring the reported result reflects the solvated state
• Raises RuntimeError if the required single point fails to converge

Test coverage (tests/test_backends_worker.py):

• Updated test_solvent_on_unsupported_calc_type_returns_error to exclude "geometry_opt" from the rejection list (now supported)
• Added test_solvent_on_geometry_opt_runs_required_final_single_point — verifies the solvated single point runs and its energy replaces the optimizer's result
• Added test_solvent_on_geometry_opt_final_single_point_must_converge — ensures unconverged solvated single points fail the job

Result serialization (quantui/results_storage.py, quantui/backends/worker_payload.py):

• Persisted cc_converged, td_converged, and n_converged_states to result.json (AUDIT F07/F08 follow-up)
• These fields were previously only in-memory; reloaded History cards had no way to distinguish "no post-HF/TD ran" from "ran but didn't converge"

Result formatting (quantui/app_formatters.py):

• Updated format_past_result() to use context-aware convergence labels: "Converged" for frequency/tddft/CC methods (where the label folds multiple convergence criteria), "SCF converged" only for plain HF/DFT
• Prevents misleading "SCF converged: No" labels when the SCF succeeded but post-HF/TD correlation failed
• Added test coverage for all three label scenarios

TD-DFT convergence logic (quantui/tddft_calc.py):

• Relaxed converged from "every root converged" to "at least one root converged" (AUDIT F08 code review)
• Rationale: Davidson solvers routinely leave higher/harder roots short of the iteration budget while lower, physically relevant ones converge — a normal, fully valid result was being flagged "treat with caution"
• Handles missing td.converged attribute gracefully (treated as "no information", not "unconverged")
• Added test coverage for partial convergence and missing attribute scenarios

Performance optimization (quantui/orbital_visualization.py):

• Cached ECP core-electron lookups per unique element instead of per atom
• Hoisted pyscf.gto import out of the atom loop
• Reduces redundant work for large molecules in cube export

Testing

• pytest -m "not network" green (full suite, exit code 0)
• ruff + black clean

Notes

• Not app-verified (Voilà/GPU/NCShare) — cloud-only session