Speed up large imports by allowing up to 5 imports to be in-flight concurrently.
The simplest approach to this would be to divide the annotations into batches
and save one at a time. However due to the variability in the time that an
individual import can take, this can lead to sub-optimal concurrency. Instead
structure the code so that we try to always keep `MAX_CONCURRENT_IMPORTS`
imports in flight at once, as long as there are that many remaining.

Fixes https://github.com/hypothesis/client/issues/5739