make delete task hammer less the metastore

currently on startup, the delete task does quite a few request to the metastore. We could optimize a bit so less calls are made, and also make it so they are not made all at once. Here are some ideas:

• DeleteTaskPipeline::spawn_pipeline could be called while holding a semaphore so there aren't 1k requests all at once.
• get_relevant_stale_splits() could be made to accept None as delete opstamp, and figure itself what is the latest opstamp and return it along, saving one call to last_delete_opstamp() per index.
• in partition_splits_by_deletes, call list_delete_tasks() only once, and sort ourselves which delete have already been applied to any given split

Start by tracing DeleteTaskPipeline::spawn_pipeline, get_relevant_stale_splits(), last_delete_opstamp(), partition_splits_by_deletes, and list_delete_tasks(). Compare the startup request pattern with the proposed semaphore and call-reduction changes; done means fewer metastore requests and no burst of all requests at once.