In many shards benchmarks we see a lot of contention when submitting tasks.
This is obvious when working with lots of large task batches and doing long iterations.

We don't need to lock in `runIfNotProcessed` past removign the task set for a key
and can be a little more efficient when it comes to creating the new tasks set in
`submitTasks` as well.
Also, we don't need to use a fully locking map as we often have operations for different
batching keys interleaved so moving to CHM as well.

This change is particularly relevant for stability because we often submit tasks from
network threads directly where grinding through a e.g. bunch of shard state updates and
having to lock on the map over and over while e.g. a huge batch of index create or so
was iterated over in `runIfNotProcessed` caused very visible latency.