LazyBackend only queried and controlled its primary communicator for errors, suspend/resume, and memory statistics. Pair communicator failures were therefore hidden, and pair allocations remained active while the wrapper reported itself suspended.

Snapshot pair communicators under the map lock, fan lifecycle operations out to each snapshot, and aggregate memory counters across all channels. Exercise suspend/resume with a real lazy P2P channel.

Test Plan:

```bash
lintrunner -a
```

Authored with an AI assistant.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/191553
Approved by: https://github.com/kapilsh, https://github.com/fduwjj
ghstack dependencies: #191510, #191511, #191517, #191528, #191542, #191544, #191549