[dundysm]

fixes #1022

MultibodyJointSet::do_insert created a root multibody and a connectivity graph node for every endpoint not yet in the set, and only then checked whether the insertion was valid. so:

• insert(body, body, ..) created two provisional roots, merged them and returned Some for a self joint;
• each rejected insert(fresh, child, ..) where child already has a parent returned None but left a root multibody and a graph node behind for fresh.

this does the checks first, from the existing rb2mb entries: identical handles are rejected, a body2 that already has a parent is rejected, and two bodies already in the same multibody are rejected. the provisional roots are only created once the insertion is known to succeed; the rest of the function is unchanged.

added crates/rapier3d/tests/issue_1022_multibody_rejected_insert.rs; on master the self joint is accepted and ten rejected inserts leave 11 multibodies instead of 1; both tests pass with the change. cargo test passes for rapier3d and rapier2d, and clippy is clean.