Before this change, we would send requests to language servers without
canceling them even if we never wait for their response.

Example: when doing document highlights, we'd send a request (modulo
debouncing) whenever a change was made, ignoring previously sent
requests that might still be in-flight.

With this change, we now send a Cancel request (from the LSP spec) to
the language server in case no one listens to the response anymore
(which is what happens when the `Future` returned by `request_internal`)
is dropped.

Release Notes:

- Improved performance when interacting with language servers.