Add a physical regression test and the diagnostics it needs

[Maccchiatooo]

CI so far only proved the solver did not crash: a NaN velocity field would
have exited 0 and passed. This adds the two exact invariants of the discrete
dynamics as assertions, plus a check that the answer does not depend on how
the domain is split across ranks.

LBM::Conserved reduces total mass, momentum and kinetic energy across all
ranks. In this pressure-based formulation sum_i f_i = 3p and sum_i f_i e_i = u,
and the equilibrium reproduces both moments exactly, so BGK collision leaves
them untouched; streaming on a periodic domain is a permutation. Mass and
momentum are therefore invariants up to rounding. Cost is one 5-double
Allreduce per reporting interval, next to an MPI_Barrier that was already
there. Output goes on its own line with no "MLUPS" token, because
tools/parse_scaling.sh selects lines with awk '/MLUPS/'.

tests/regression.sh runs 32^3 for 20 steps on 1, 2 and 4 ranks and asserts:
mass drift < 1e-12 relative, |net momentum| < 1e-10, and agreement between
rank counts < 1e-12. Tolerances are ~100x the measured values (1.2e-14,
5.1e-14, 1.4e-14 on Kokkos 5.1.1 OpenMP + Open MPI).

Kinetic-energy monotonicity is deliberately not asserted: the distributions
start at equilibrium with an unrelaxed pressure field, so the opening steps
carry an acoustic transient and ke really does oscillate (9.852 -> 10.219 ->
9.889). That assertion would have been flaky and wrong about the physics.

Verified the test can fail: perturbing pp by a relative 1e-12 in Collision
drives the mass drift to 3.8e-11 and turns all three rank counts red.