[Feature Request] Barrier and memory fence operations

1. Implement a barrier mechanism and support efficient barrier operations, including:
- `T.create_barrier(num_arrive=num_arrive)`
- `T.arrive(barrier)`
- `T.wait(barrier)`
- `T.sync(barrier)`

so that it can be used to do p2p synchronization or group synchronization.

2. Implement memory fence to ensure correct memory ordering and access semantics.