Document thread-safety issues in C callbacks

[timholy]

The main reason I made this a PR is to ask: should we export SingleAsyncWork? The only other place I see the manual making use of an unexported function/type is in networking-and-streams.rst, which uses Base.getaddrinfo.

On balance I suspect things are fine as they are, just thought I should raise the issue.

[JeffBezanson]

This is great documentation to have --- unfortunately highlighting how ugly the current available workaround is :)

@loladiro is this writeup correct?

getaddrinfo is now exported.

[Keno]

Yes, this looks correct.

[timholy]

The fact that this is correct is entirely due to @vtjnash, who on julia-users provided the clue to use uv_async_send.

Should we merge this, or should we wait for a cleaner interface? At the cost of an additional export, I could probably write a function (which would have an eval or two) that would simplify this a bit. But it would be pretty limited: the user-supplied callback would have to be a thunk. That means that even if the C callback supplies some data, it couldn't get to the Julia callback. (That's officially true now, but if you are willing to experiment with danger I'm guessing you could stash stuff in a pre-allocated global buffer before calling uv_async_send.)

I dug into this in the context of needing a "wakeup" callback, so for my usage a thunk is fine. But I wonder if it would be too limiting to deserve a nice wrapper, and that at this point we're better off just documenting an ugly interface.