Spans lifetime management

[crusaderky]

• Follow-up to User-defined spans #7860

Spans currently live forever on the scheduler, and with them the TaskGroups they aggregate.
This meta-data outlasts the Tasks, which are instead dereferenced as soon as they are forgotten, and will slowly increase memory usage on the scheduler.

For clusters that last a few hours, this should not be a problem, as the number of TaskGroups does not scale with the number of tasks: a workflow running on 1 megabyte of data will generate the same amount of permanent metadata as the same algorithm processing 1 petabyte.

In situations where the scheduler is never restarted, this may become a problem.
Ensure that spans are dereferenced after a certain amount of time (if and only if all of their TaskGroups are already been forgotten by the scheduler!), and with them their TaskGroups.

Note: TaskGroups are also referenced by Computations, which is deprecated functionality conceptually similar to spans, and that the Spans replaced.

[fjetter]

Unless they take up any significant memory, I don't consider cleaning them up very low priority