Allow `PrimitiveJob` to use Python processes rather than threads in `_submit`.

[kshyatt-aws]

What should we add?

In primitive_job.py:

def _submit(self):
        if self._future is not None:
            raise JobError("Primitive job has been submitted already.")

        executor = ThreadPoolExecutor(max_workers=1)  # pylint: disable=consider-using-with
        self._future = executor.submit(self._function, *self._args, **self._kwargs)
        executor.shutdown(wait=False)

Any backend that's using PrimitiveJob is locked into using Python's threading, which can cause locks when interacting with other non-Python languages with their own thread schedulers. Python processes don't have this problem (because they avoid the GIL). This behaviour is currently blocking me when using a custom backend I've written. It'd be great if the PrimitiveJob could obey the preference set in the QISKIT_PARALLEL environment variable.

[kshyatt-aws]

Another option which might unblock me would be to allow the user to supply an initializer option to the ThreadPoolExecutor

[t-imamichi]

It makes sense to control the thread pool, e.g., do not use ThreadPoolExecutor if QISKIT_IN_PARALLEL is True.
It is related to #12963. Since PrimitiveJob currently stores the result and status in the ThreadPoolExecutor, we need to store them directly in PrimitiveJob.