Prismatic axis rows for rotation-locked pairs; a spring of no stiffness lets go

[nicolas-maman]

Fixes two defects inherited from upstream, found while auditing the joint solvers against Box3D e77352c. The rest of the per-step math matches upstream line for line.

1. Prismatic spring, motor and limits skipped for rotation-locked pairs. These ran only when the pair could turn (fixed_rotation == false). They are linear rows along the axis, and their mass comes from the bodies' masses, so they work whether or not the bodies can rotate. A rotation-locked body on a prismatic joint (an elevator or a piston) got no motor drive and no spring pull, and slid straight through its limits. Only the rotation lock keeps that gate now; the wheel joint's linear rows never had it.

2. A spring set to zero stiffness kept pushing. make_soft(0) gives no mass and no impulse scale, so the solve's delta was zero and the last spring impulse stayed. It was then warm-started every sub-step for as long as the joint lived, so a ragdoll joint that was let go kept pushing. The port already drops such impulses for the distance and motor joints. The prepares now drop them too when the spring is off or has zero stiffness, for:

• the revolute, prismatic and spherical springs;
• the wheel's suspension and steering springs;
• the parallel joint's soft row.

test_physics_world's test_joint_springs_and_locks:

• a rotation-locked elevator stops at its limit; on main it falls to y = -18.
• a hinge spinning against a spring set to 0 Hz keeps spinning without being pushed; on main the frozen impulse spins it up from 0.25 to 21.8 rad/s.

All 28 test programs pass locally on aether 0.772.0, with no warnings.