Add subgradient (augmented-Lagrangian) penalty update strategy

[bernalde]

Motivation

The iterative penalty-refinement engine (JuliaQUBO/QUBO.jl#67, design record in this comment) ships with multiplicative escalation for MQT QAO parity. Multiplicative escalation needs ρ→∞ for feasibility and collides with sampler coefficient dynamic-range limits within a few ×10 steps (Ohzeki, Sci. Rep. 10:3126, 2020, names this explicitly). The method of multipliers converges with bounded ρ (Powell's shifted-penalty view; Bertsekas 1982: linear rate at fixed finite ρ), and Yonaga–Miyama–Ohzeki (arXiv:2012.06119) run exactly this iteration on a D-Wave annealer with fixed ρ = 0.1, tracking best-feasible and best-energy samples separately.

Proposed change

Building on the engine from #67's PR A:

• Attributes.AugmentedLagrangianPenalty{T} <: ConstraintPenaltyMethod with multiplier::T (λ, any sign, default 0) and rho::T (> 0): emits λ·ℓ + ρ·ℓ² for equalities and the sign-oriented composite for LessThan/GreaterThan, reusing _unbalanced_penalty's emission (src/compiler/constraints.jl:61-71) but lifting UnbalancedPenalty's strictly-positive-coefficients and inequality-only restrictions. Document as iterated unbalanced penalization for inequalities (the Rockafellar one-sided AL term is not QUBO-expressible slack-free).
• SubgradientUpdate strategy for Attributes.PenaltyUpdateStrategy: per iteration, from the best sample's signed residuals (_constraint_measurements; EqualTo residual is signed, inequalities positive-when-violated): equality λ ← λ + η·r; inequality λ ← max(0, λ + η·r); default step η = ρ (classical MoM). Safeguard: escalate ρ ×τ only when ‖r_k‖ > γ·‖r_{k−1}‖ (Bertsekas γ ≈ 0.25, BCL-style), so escalations stay rare and coefficients bounded.
• MAX-sense handling via the compiler's σ convention (multipliers live in the minimization frame; tests must cover a MAX model).
• Interval constraints: the measured residual loses violation direction (feasibility.jl:357-367) — either split handling or document exclusion.

Acceptance criteria

• An under-penalized equality-constrained model reaches feasibility via multiplier updates with ρ held fixed (no escalation), within the iteration budget.
• An inequality model converges with one-sided updates; multipliers never go negative.
• A MAX-sense model converges with correct signs.
• Safeguarded ρ escalation triggers only on stalled residuals (unit test on the trigger condition).
• Multipliers and iteration count queryable post-solve; docs section with the references above.

Part of the MQT QAO parity epic: JuliaQUBO/QUBO.jl#69 (extends beyond parity — this strategy has no MQT equivalent).