Unify the polished-7R solver tail into refinement.polish_candidates

[siddhss5]

Problem

Three solvers hand-roll the same polish tail, and it has already started to drift (the #462 completeness invariant -- filter limits after polish, not before -- currently lives in only one of them):

• seven_r/srs_polished.py (solve)
• seven_r/spherical_shoulder_polished.py (_polished)
• seven_r/_swivel_limits.py (resolve_in_limits, approximate path)

Each builds identical _fk/_jac closures around poe_forward_kinematics/kinbody_jacobian, calls lm_refine_batch, keeps rows with res <= atol (and, for two of them, in-limits), then dedup_by_wrap_close. ~20 lines x 3, plus the invariant duplicated.

This sits one layer below postprocess.finalize_solutions (which is Solutions -> limits -> seed-rank -> truncate); the new helper is raw seeds -> polish -> FK-accept -> dedup -> Solutions.

Plan

1. Extract refinement.polish_candidates(kb, seeds, T, *, accept_fk_atol, dedup_tol, lm_fk_atol, lm_max_iters, limits=None, batched=False, solution_factory=None, max_solutions=None) -> list[Solution]. The completeness invariant (polish first, filter limits second, never pre-filter) is documented and enforced here, once.
2. Delegate all three paths to it, preserving each caller's exact params (behavior-preserving; validated against the existing green SRS/7R suite). Callers keep only their arm-specific seed generation.
3. jointlock.seven_r is deliberately out of scope: exact HP per locked-joint sample, no polish, in-sweep filter on exact configs. Its gap (ROBOTERA Star1 7R: respect_limits branch-coverage gap (tier-1 jointlock+HP), raw solver is complete #459) is sampling density, a different fix.

Payoff

• Booster T1 7R: ~5% coverage gap in seven_r.spherical_shoulder_polished on a 70mm-offset wrist #460 (Booster / spherical_shoulder_polished) collapses to a one-line change: feed unfiltered dense seeds; the shared helper guarantees filter-after-polish.
• A future polished-solver can't reintroduce the pre-filter bug.

Follow-up to #462 / #465.