Allow overriding parallelism globally for ParallelCompile

[thesamesam]

Hi,

ParallelCompile currently defaults to the number of available CPUs with no override unless consumers explicitly pass an environment variable of their own to allow choosing it. This means for a distribution, there's no way to globally constrain pybind11-generated threads for compilation.

We've had some weaker machines become overloaded by this when testing in Gentoo.

Would a patch be welcome to accept some fallback envvar that is respected if the named one from a consumer isn't set?