feat(client): add Channel::shutdown() to terminate the task explicitly

[mdmzfzl]

Problem

The client task completes only when every Channel clone has been dropped:
Receiver::recv() returns None once the last sender goes away, which becomes
SessionError::Shutdown.

That makes termination depend on how many handles exist and where they are held.
A caller that publishes a Channel somewhere outliving the task — a service
registry, a shared cache, anything with process lifetime — can never terminate
it, and awaiting the task's JoinHandle hangs.

Dropping handles in the right order is a convention every holder has to honour,
and nothing in the type system enforces it.

Change

Adds Channel::shutdown(), which sends shutdown as a command rather than
signalling it by releasing the last handle:

pub async fn shutdown(&self) -> Result<(), Shutdown>

Drop-based termination is unchanged, so this is additive — existing callers need
no changes.

Command gains a Shutdown variant, handled at both points where
reads its queue:

• run_cmd, while connected, returning SessionError::Shutdown
• fail_next_request, while disconnected or waiting to retry, r
  StateChange::Shutdown

Both already existed for the drop case, so this reuses the wind-down path rather
than adding a second one. Handling only the first would leave
solely while the device is reachable.

Semantics

• Queued behind pending requests, so anything already in the queue is still
  processed — matching what dropping every handle does, since an
  buffer before reporting closure.
• A transaction already in flight runs to completion, so the t
  within one response timeout of the last queued request.
• Requests made after shutdown fail with RequestError::Shutdow
• Signals the task rather than waiting on it; await ClientTask::run to observe
  completion.
• Returns Err(Shutdown) if the task has already terminated, so a second call is
  harmless.

TCP and serial share ClientLoop, so both are covered with no
channel task.