[DNM] attempt to use a semaphore to work around opte 758

[sunshowers]

This is an incomplete attempt to use a semaphore with one permit, to prevent kstat reads and XDE ioctls from happening simultaneously, with the goal of working around oxidecomputer/opte#758 (see #9211).

What's missing:

• Tests need to be updated.
• There's no great way to ensure that every ioctl is covered -- would need manual review to ensure that it is.
• Dropping an illumos_utils::opte::PortInner (a Port is an Arc around this) calls delete_xde, which would need to be protected by the same semaphore.
• Validating that this actually works, doesn't introduce new issues, etc.

---

Unfortunately this is becoming complicated and fragile enough that I've become significantly less optimistic about using this path as a workaround, compared to the alternative where:

• support monitors this as part of the r17 upgrade, and
• we fix it in the illumos kernel for r18

In particular, the requirement to pass the semaphore into the PortInner and lock it inside the generally-unobservable Drop impl has me quite concerned.

Putting this up there for folks to look at in the meantime.

[jclulow]

FWIW, it might be worth using DTrace to collect user stacks from ioctl system calls made by sled agent during some test runs.