The error(s) gets recorded at failure time, and the list is printed at
job deletion time.

This allows helgrind to be aware of succession of events.
Also reduced the default queue depth for unit tests.

Adapted the in-RPM script generation.

The delay will initially slow down the user and then the semaphore will block.
This prevent the queue from going too big and the agent object from growing
so big all becomes slow.

Better to have all formatting on the client side, so we format the
header and the reponse values in the same place. (Also no need to
send this data over the wire as it's always the same).

PB now accommodates "lpa" and "lpr" responses as well as "af ls"

The test relies on this order.

The new default is 10 for unit tests.
The number is set by the frontend, which is the only process where it matters. The default
in the configuration file is set to 500 for CI tests.

The requeust is now created first so the data is safe before returning to the caller.
Also improved the timing logging.

This version is not data safe as the request is created late, right after
referencing the job to the queue.

The size is set to 500 to allow unit tests to pass.
Next steps are: port the mechanism to retrieve, and make the thread pool size configurable.

The top half goes as far as when the request is safe in the object store.
At that point, the bottom half is launched in a new thread, and success is returned
the called. This will enable lower latency for the users while retaining
the same data safety.

This version is experimental aas it spans an undetermined number of threads. A more controlled
version with a work queue should be implenented on the long run.