Struct Tp2ReplicatedRowJoin 

pub struct Tp2ReplicatedRowJoin { /* private fields */ }

Persistent two-rank all-reduce for one replicated f32 row.

Each rank pushes its local partial directly into a peer-resident staging row, records one reusable event, waits for the peer’s corresponding event, then adds (rank0, rank1) in that same operand order on both devices. Two staging rows per direction are enough because join j + 1 is ordered after each rank’s add at join j, so overwriting parity j cannot race its peer consumer. The collective allocates nothing and performs no host synchronization per call.