Improve lifecycle management of LSP

[DavisVaughan]

@lionel- and I were looking at the LSP lifecycle a little more and realized that when the tower-lsp server exits here:

ark/crates/ark/src/lsp/backend.rs

Line 427 in b8505c5

server.serve(service).await;

we should also unset the events channel in RMain that we sent it here:

ark/crates/ark/src/lsp/backend.rs

Lines 397 to 403 in b8505c5

	r_task({
	let events_tx = events_tx.clone();
	move || {
	let main = RMain::get_mut();
	main.set_lsp_channel(events_tx);
	}
	});

because there is no longer anything it can send messages to.

The tower-lsp server will exit like this during a Positron developer refresh, so it is a decently common case. I think it is possible for us to get in a bad place if we don't clear the RMain sender.

[lionel-]

See discussion in #617 (comment)

[lionel-]

If we solve the issue of zombie extension hosts, it would also be good to detect whether an LSP is already running on comm_open. That would be an error as the client should first shut down the LSP.

Unfortunately comm_open is a notification, not a request, so there is no straightforward way of propagating the error. We might want to extend this message and allow it to be a request if a field is set. This would be advertised as a kernel capability and allow the client to wait for a result and know for sure whether an LSP was started.