And display delayed message counts and delayed retry configurations.

`DiscBytes` stands for "discarded bytes" which was for the old field
name `discarded_bytes` before it got renamed to `reclaimable_bytes`.

Previously, when a quorum queue consumer timed out, RabbitMQ detached
the AMQP 1.0 link with an amqp:resource-limit-exceeded error. This
is a harsh response: the client loses the link and must re-attach to
continue consuming.

Instead, follow the AMQP 1.0 spec [2.6.12] which allows the source
to spontaneously transition a delivery to a terminal state. When a
consumer times out, RabbitMQ now sends DISPOSITION(role=sender,
state=released, settled=false) for the timed-out deliveries. The
client is expected to confirm by sending DISPOSITION(role=receiver,
settled=true) back.

The Erlang AMQP 1.0 client (amqp10_client) is updated to handle
this new DISPOSITION(role=sender) frame: it removes the affected
delivery IDs from the link's incoming_unsettled map, auto-settles
by replying with DISPOSITION(role=receiver, settled=true) and notifies
the application.

On the broker side, when the client settles back with the released
outcome, the normal requeue path (rabbit_queue_type:settle/5 with
requeue op) is followed. To make this work end-to-end, rabbit_fifo's
return() function now calls maybe_untimeout/2, clearing the
consumer's timeout state so it can receive messages again. Previously
maybe_untimeout was only called from complete_and_checkout (the
settle/complete path), meaning consumers could never recover from a
timeout via the requeue/return path.

The integration test verifies the full flow:
1. Messages are sent and received but not settled
2. Consumer timeout fires, client receives the released disposition
3. Client lib auto-settles, consumer timeout state is cleared
4. Timed-out messages are redelivered, new messages are received
5. Redelivered messages have correct headers (delivery_count=0,
   first_acquirer=false, x-opt-acquired-count=1)

Log a warning once per outgoing link when a consumer timeout fires,
including the connection name, link name, handle, AMQP container ID,
and queue name. The warning is emitted only on the first timeout for
each link to avoid log noise.

Expose the consumer timeout state in the Management UI by adding a
"Consumer timeout" boolean column to the outgoing links table. The
cell is highlighted with a yellow background when true, consistent
with how other notable conditions (e.g. zero link credit) are
displayed. A help icon explains the field.

x-acquired-count is more symmetry to x-delivery-count.
(Renaming x-delivery-count would be a breaking change.)

This is for consistency with the other stashed_* fields instead and not
really for performance.

Snapshot frequency was too high for shallow, fast-flowing queues
because the reclaimable-bytes threshold was proportional to the
(small) approximate snapshot size. With many queues sharing a WAL,
this resulted in multiple snapshots per queue per WAL cycle.

Scale the minimum reclaimable threshold inversely with the WAL fill
ratio so that snapshots are deferred until the WAL is sufficiently
full, yielding roughly one snapshot per queue per WAL cycle.

`gb_trees:size/1` is constant time - the size is stored at the top-level
tuple.

    -doc "Returns the number of nodes in `Tree`.".
    -spec size(Tree) -> non_neg_integer() when
          Tree :: tree().
    size({Size, _}) when is_integer(Size), Size >= 0 ->
        Size.

So we don't need to manually track the tree size with a len field. This
change removes the field.

rabbit_fifo: Use optimised tuple for delayed keys