Race condition in trio.hazmat.WaitForSingleObject

[njsmith]

https://github.com/python-trio/trio/pull/575/files#diff-ed34813c30595dc83bb2a66e0973e22dR54

This appears to have a race condition. The new thread might start waiting only after the handle has been closed. This could be fixed by making the thread close the cancel_handle.

Originally posted by @GSPP in #233 (comment)

[njsmith]

Hmm, I think you're right. Probably difficult to hit, but not impossible.

I don't think the simple fix is quite right either though – if we move the CloseHandle into the thread, then we introduce the possibility that the handle will already be closed before we try to SetEvent in the finally block.

I guess we could use some kind of reference counting scheme, where whichever of the child thread/main thread exits last is the one who calls CloseHandle?