Change algorithm of tootctl search deploy to improve performance

[Gargron]

Move importing logic into index-specific classes. This allows using custom logic depending on index. For example, for the statuses index, since statuses are only searchable under specific conditions, it is quicker to iterate over those conditions than over the entire statuses table. Use quick estimates for the workload instead of running an exact COUNT query on each index. This takes a long time even before real work begins and in the end, the count will not stay accurate anyway. Disable refreshing on indexes that are about to be imported to increase importing speed, then re-enable afterwards. Add a step that iterates over documents in the index and finds the ones that no longer exist in the database, to remove bloat. Those things re-add functionality that was already present in Chewy's own Rake tasks, but Chewy code is not appropriate for large data sets (for example, it loads all document IDs into a single array before checking if they still exist).

The concurrency-related logic is once again changed. Instead of splitting a single batch into slices and concurrently processing those slices (and waiting for them to be processed before moving to the next batch from the database), each batch is queued into the thread pool, which has a maximum queue size. When the queue is full, new batches have to wait before being queued (backpressure). The batch size parameter therefore now transparently corresponds to the amount of documents that get imported into Elasticsearch in a single bulk request.

There was also a bug that prevented the desired skipping of indexing of statuses with empty searchable_by fields due to the wrong key name (to_index (wrong) instead of index (correct)).

Other changes include:

• Use tag history's aggregate method which is faster than using reduce
• Use cached following/follower counts instead of counting local followers for accounts