Calling `Scheduler.activate_monitoring()` from the `Scheduler`
constructor didn't have any effect because `self.activate_monitoring()`
always exited early.

This change fixes the boolean check before that function call so that
the monitoring structures get initialized properly.

Try fetching the `CacheCapabilities.max_batch_total_size_bytes` value
from the server and, if the server provides a limit, respect it for
batch transactions.

Also, make the threshold value for determining when to queue a file
for fetching in a batch request a configurable % of the maximum
batch request size.

(Addresses #100)

#160

#160

#160

#160

#160

#160

#160

This should prevent a race between delete_job_operation() and
delete_job_lease() when dropping Job objects after execution.

#160