feat(broadcast): add unbounded policy

[orthur2]

This PR implements the unbounded policy proposed in #95. Messages remain in the shared buffer until every active receiver has consumed them, or until receivers that have not consumed them are dropped. The channel has no capacity limit, so receivers do not lag and the receive error type has no Lagged variant.

The WaitSet cleanup needed by Recv::Drop was handled separately in #116. This PR is based on that change and focuses on the unbounded policy.

The channel uses a single Mutex to protect the VecDeque<Arc> buffer and the Slab containing receiver cursors. It also tracks how many receivers are at the current buffer head. When the last receiver at the head advances or is dropped, the channel finds the new minimum cursor and removes the corresponding messages. Removed values are dropped after releasing the lock.

Messages are stored as Arc internally, while recv returns a cloned T, so callers do not need to use Arc themselves. subscribe and resubscribe both create a receiver at the current tail without changing the backlog of existing receivers. Message versions use u64, and send panics if the version counter overflows.

By the way, I will update this branch if the WaitSet design changes.

Refs #95. Based on #116.