fix(qp): align fast/cvxopt objective convention, reject dropped equalities, tune PDIPM

[bardhh]

Correctness

• fast and cvxopt backends solved the wrong objective. The registry convention is min x'Hx + f'x, but both passed H straight into their native 0.5 x'Px + q'x form, tracking 2·u_nom instead of u_nom. Hidden in safety tests because the active barrier pins the solution. Fixed with P = 2H at the registry boundary; all four backends now agree on the audit probe (u_nom=[1.0,-0.5] → [1.0,-0.5]) and on a new cross-solver parity suite (interior / single-active / corner / random / end-to-end CBF-filter cases).
• fast silently dropped equality constraints, so MPC through get_solver("fast") returned dynamics-violating trajectories with success status. Now raises NotImplementedError pointing at jaxopt/casadi.
• test_fast_qp.py expected [2,4] — the bug's answer. The correct minimizer of x'x − 2x₀ − 4x₁ under this convention is [1,2].

PDIPM performance (1.64× per solve, bit-identical results)

• Factor the reduced Newton matrix once per Mehrotra iteration (predictor and corrector differ only in RHS); residuals computed once for the convergence check and the step.
• DEFAULT_MAX_ITER 25 → 16, calibrated on 250 benign + 250 ill-conditioned near-boundary QPs: deviation vs max_iter=60 is exactly 0.0 on both families. (12 fails accuracy on 30/250 hard QPs precisely when the barrier is nearly violated — table in qp_solver_pdipm.py.)
• fori_loop retained deliberately: while_loop has no reverse-mode AD rule and the differentiable-CBF-QP work takes gradients through this solver. The max_iter=25 pins in solve_qp_fast/fast_solver now defer to the solver default.

Verification

• 527 passed, 5 skipped on this branch (baseline was 492).
• New tests fail against pre-fix code: 21 failures when the source is reverted (parity, guard, controller regressions), so they pin the fix rather than passing vacuously.
• Reverse-mode jax.grad through the solver verified finite.

Merge order: this is PR 1 of a 4-PR stack (merge top-down).