Subactors don't de-register from arbiter on cancel?

[goodboy]

I noticed while working on piker that we are leaking registry entries when a subactor is cancelled.
Need tests and a fix for this stat!

[goodboy]

Figured it out: it's a regression from #128 which is actually my fault.
Turns out if you __aexit__() a trio.Process it'll kill the process on cancellation before actor teardown happens (which is not the "soft kill" we want).

Gonna add some tests for this now that I think I've got a fix.

[goodboy]

Huh, on even further investigation it's only when subactors are cancelled and the root actor is not the arbiter that the problem arises; super interesting.