[core][compiled graphs] Don't require finishing previous executions in the case that Python destructs dagrefs out of order

[dayshah]

Description

With #49781, we fixed the case where the destruction of a later dagref will result in the invalidation of previous dagrefs from previous executions. However, that requires us to wait until all previous executions are finished before we can finish the del call for that dagref.

This can becomes an issue because Python doesn't guarantee the order of destruction so if you have a script like

ref1 = compiled_graph.execute(1)
ref2 = compiled_graph.execute(2)
# both are supposed to be destructed somewhere here

Python could call del on ref2 before ref1 and then we would wait for the first execution to finish even though we're not using the result from either.

This shouldn't be a common case, but something that we should think about solving for if we start seeing it happen more often.

Use case

No response