test_runner: do not reuse a worker ID held by a running file - #65739

WorkerIdPool handed out worker IDs round-robin ((nextId++ % maxConcurrency) + 1)
and never released them, so an ID was only unique among the first N test files.
As soon as files finished out of order, a file that started later was given an ID
still held by a live process, which defeats the purpose of context.workerId:
allocating a database, port or directory per worker.

With --test-concurrency=2 and three files, where the first one outlives the
second:

start slow 1
start fast 2
end fast 2
start last 1   <-- same ID as the still-running 'slow'
end last 1
end slow 1

Track the IDs that are actually in use and hand out the lowest free one, then
release it once the child process is gone. The release happens in a finally
block so an aborted run or a failed spawn does not leak the ID out of the pool.

The pool no longer needs to be told the concurrency level. That also removes a
mismatch: the pool was sized from globalOptions.concurrency ?? concurrency,
which can differ from the concurrency the root test actually enforces. Since IDs
are handed out lowest-first, the highest ID in use is now bounded by how many
files really run at once.

Refs: #61394