In Qiskit#5324 we changed the default for running with parallel map on macOS
to default to false unconditionally, but that was likely too strict.
While there are a myriad of issues related to Python multiproccessing
and macOS it doesn't fail in every macOS environment. Additionally the
performance benefits from running in parallel is signtificant. This
commit attempts to make the default change a bit more nuanced. Instead
of defaulting to false on all macOS systems it only does that for Python
3.8 and macOS. This is because in python 3.8 the use of spawn instead of
fork means in the best cases if it works completely as expected it will
be slower (see $5323) but also means there are unforseen consequences
(see Qiskit#5376) because it's spawning a new interpreter instead of forking.
But, often it just doesn't work. For other python versions there are
still potential issues on newer versions of macOS (>=10.14) so to catch
potential issues there a small parallel benchmark is run on import to
ensure that it both works and is faster than serial. If both are true
then we default it parallelism to True, otherwise it's False.

Fixes Qiskit#5373

The use of the parallel benchmark for macOS to determine whether
parallel execution is actually faster or not was hanging on import. I
expect this was due to it being executed on import and not having
sufficient context awareness in the subprocesses to handle that. To
avoid the hang this just makes the default for macOS parallel processing
to be True if Python < 3.8.