fix(lsmr): keep MGK's ⟨v, Mv⟩ in range at extreme ‖A‖

[schroedk]

Stack 4/5 (extreme scales), on main after #403.

• MGK normalizes u in place once β leaves 1e±64, since Aᵀu scales with ‖A‖β; inside that band the stream is unchanged.
• alpha_from_vp lets a NaN ⟨v, Mv⟩ (∞ − ∞ from finite terms) reach its re-sum, which scales both norms to 2^500 so a cosine below 2^-1074 survives; a subnormal negative ⟨v, Mv⟩ takes the re-sum's sign.
• A non-finite solution x fails with InvalidInput. Past Krylov exhaustion at tol = 0, h̄ overflows and x became NaN, which budget stops returned as Ok (plain lsmr on main too; mlsmr once this PR recovered the overflowing dot).

Preconditioned solves are exact up to raw ‖A‖ ≈ 1e244 with ‖M⁻¹‖‖A‖ finite; past it they can fail with InvalidInput. The power-of-two scaling sweep gains an mlsmr arm.

Mutants: never normalizing u fails the sweep and the overflowing-dot test; always normalizing fails the headroom pin; the old re-sum fails both alpha_from_vp unit tests; is_finite for is_normal fails the subnormal-pair test; dropping the finite-x guard fails both exhaustion cases.