mimetypes.init is slow on Windows because of the big registry search, which involves touching thousands of entries in order to add a few hundred into the initial table.

Things get even worse when audit hooks are enabled, because the winreg methods need to be hooked. However, we know that this particular operation is coming from core, and so we could skip these hooks.

Both issues can be trivially solved (helped) with a native accelerator function.

It's only a rough benchmark, but my timings for `python -c "import mimetypes; mimetypes.init()` were ~200ms without the accelerator and ~100ms with the accelerator.

Since that includes all of startup, it's actually a better than 50% reduction for mimetypes.init(). I'm happy to take that without questioning precisely how much better it is ;)

mimetypes doesn't _really_ support clearing, but `python.exe -m timeit "import mimetypes as M; M.types_map.clear(); M.init()"` is pretty close.

Without the accelerator: 133ms
With the accelerator: 27.8ms

New changeset bbf2fb6c7ae78f40483606f467739a58cd747270 by Steve Dower in branch 'main':
bpo-44582: Accelerate mimetypes.init on Windows with a native accelerator (GH-27059)
https://github.com/python/cpython/commit/bbf2fb6c7ae78f40483606f467739a58cd747270

New changeset 08697ac5d1543fca3629c719ab43e50d73021631 by Miss Islington (bot) in branch '3.10':
bpo-44582: Accelerate mimetypes.init on Windows with a native accelerator (GH-27059)
https://github.com/python/cpython/commit/08697ac5d1543fca3629c719ab43e50d73021631

Dropping the 3.9 backport. Pretty sure it only needs a clinic regen, but it's also not worth the risk of changed behaviour.

We could consider it again after 3.10 has shipped if people really want it (and can help validate it).