Soundness bugfix for barrier<thread_scope_block> on sm_70

[gonzalobg]

For sm_70, barrier arrive has an optimization to "coalesce" all arrives with the same update to the same barrier into a single update performed by a "leader" thread.

This optimization is missing a release fence to establish cummulativity between all coalesced threads and the leader, before the leader performs the update.

[gonzalobg]

@daniellustig could maybe review?

[daniellustig]

The cumulativity fix seems reasonable to me

[gonzalobg]

@wmaxey ?

[wmaxey]

Thanks for review, David. Merging.