Use better http server than httpuv in curl_echo

[jeroen]

Because httpuv now is always daemonized via later it does not synchronize very well anymore in curl_echo. So we run into these sorts of problems: #327

Perhaps we can create something simpler.

[shikokuchuo]

NNG provides an http server implementation, which is currently not exposed by nanonext.

A demo exists e.g. at r-lib/nanonext#48 - but if it's just an echo then it should be easy to implement on a thread in-process and it would run independently of the R event loop as it wouldn't touch any R objects.

[jeroen]

The easiest solution would be a simple non-threaded http server, that runs in the foreground but without blocking. Previously this was possible in httpuv and we would process the messages queue by calling httpuv::service() and it would return when done. So we could alternate calling httpuv::service() with doing curl work, without threading problems.

[jeroen]

For now I have an alternative solution curl 6.2.1 (on cran): https://github.com/jeroen/curl/blob/master/src/dryrun.c

If we somehow want to do this without httpuv, I think the best approach is bundling libhttpparser and run that on a local socket, and then we use the on_message_complete callback to abort the request and return the buffer.

[shikokuchuo]

With an NNG solution, you'd be able to synchronize, so there wouldn't be problems just because it's threaded. nanonext already wraps synchronization primitives, so it'd be easy to use them in your testing code to wait for http server events.

Just making you aware of this possibility, as it strikes me as less of a 'workaround'.

[jeroen]

If you want to do a PR I'll review it. Given that this is mostly for debugging purposes it's not super high priority to optimize it.