[Bugfix] Misc bug fixes

[majunze2001]

• Fix task refcount for sharing
• Fix typos and unify example apps
• Fix GPU allocation for TP task executors

[gemini-code-assist]

Code Review

This pull request updates example models, re-enables actual GPU allocation in the resource manager, and implements a mechanism to allocate GPUs with consecutive local ranks to support tensor parallelism. It also refactors task registration to ensure that the cluster-wide refcount is bumped for all shared tasks. The review feedback highlights two critical issues: first, skipping the refcount bump when an instance name is not yet committed can lead to premature teardown of shared tasks, which should be resolved by polling and waiting; second, raising an allocation error immediately on the first node that lacks consecutive free GPUs causes premature failure, and the allocator should instead continue searching other nodes in the priority queue.

[gemini-code-assist]

If the instance name is not yet committed because a concurrent deployment is still in Phase 2, skipping the refcount bump entirely will cause the cluster-wide refcount to be undercounted. This can lead to premature teardown of the shared task when one of the active users unregisters it. Instead of skipping, we should poll and wait for the concurrent deployment to commit the instance name.

Suggested change

	instance_name = self.unit_task_instance_names.get(task_id)
	if instance_name is None:
	# The matched task is still being deployed by a concurrent
	# declare_used and has not committed its instance name yet
	# (Phase 3). Skip the bump rather than crash — this is the same
	# rare window the rollback path below also tolerates with .get().
	logger.warning(
	"Skipping CR refcount bump for task %s: instance name not yet committed",
	task_id,
	)
	continue
	instance_name = None
	for _ in range(100): # Wait up to 10 seconds (100 * 0.1s)
	instance_name = self.unit_task_instance_names.get(task_id)
	if instance_name is not None:
	break
	await asyncio.sleep(0.1)
	if instance_name is None:
	logger.warning(
	"Skipping CR refcount bump for task %s: instance name not yet committed after timeout",
	task_id,
	)
	continue

[gemini-code-assist]

Raising CannotColocateError immediately when the first node in node_priority does not have consecutive free GPUs will cause the allocation to fail, even if there are other nodes in the priority queue that have consecutive free GPUs. Instead of failing immediately, we should continue to the next node in node_priority and only raise the error if we have exhausted all available nodes.

Suggested change

	if len(run) != num_gpus:
	raise CannotColocateError(
	f"Cannot allocate {num_gpus} GPUs with consecutive local ranks on a "
	f"single node. Free GPUs on {node} are too fragmented.",
	)
	if len(run) != num_gpus:
	if not node_priority:
	raise CannotColocateError(
	f"Cannot allocate {num_gpus} GPUs with consecutive local ranks on a "
	f"single node. Free GPUs on {node} are too fragmented.",
	)
	continue