Remove necessity for `fork` process-start mechanism in multiprocess benchmarks

[wence-]

Per #597, Python 3.14 changed the default start mechanism for multiprocessing to use "forkserver" rather than "fork". This was a safety thing (using fork correctly is very hard).

I am not sure that it is necessary for the benchmarks to start the subprocesses with fork, or if (because that was the default) we just picked up child processes that inherited the parent state and hence could use "implicit" data sharing.

In any case, ideally we probably wouldn't require fork as the multiprocessing mechanism.