Fixes race condition when epoll_ctl() gets called after the socket has
been closed, which may affect a different socket created by another
thread meanwhile.

When rpc_reconnect_requeue() gets called from inside nfs_service(),
the NfsInputStream can stall completely because the old socket has
been unregistered from epoll automatically, but the new one has never
been registered.  Therefore, nfs_service() will never be called again.

This kludge attempts to detect this condition by checking
nfs_which_events()==POLLOUT.

https://bugs.musicpd.org/view.php?id=4081

Make sure we're using the same EventLoop as the NfsManager.

Eliminate dependency on io_thread_get().

With "ok==false", newly created threads may quit instantly.

libupnp provides seconds, not whatever time unit is used by
std::chrono::steady_clock.

libcurl's WRITEFUNCTION is pretty fragile; if we destroy the CURL*
instance or even unregister it using curl_multi_remove_handle(),
libcurl will crash instantly.  But still we need to be able to handle
exceptions from inside the WRITEFUNCTION, and call
CurlResponseHandler::OnError(), which may destroy the whole thing.  As
a workaround, I use DeferredMonitor to postpone the OnError() call
into a stack frame which is allowed to destroy the request.

Let the caller decide what to do with the exception.

When the request is done, only unregister the CURL* handle, but do not
delete it yet - it may still be needed for CURLINFO_RESPONSE_CODE.

Convert assertion to runtime check.  This is useful because this is a
public method, and the caller has no chance to check if the object is
still registered.

Fixes a crash that can occur due to recursion from InvalidateSockets()
to ReadInfo() to CurlRequest callbacks.

This allows constructing an instance in any thread, and register it
inside the IOThread later.

Eliminate one unnecessary dependency for debug programs which don't
need the InputStream API.

The CurlRequest gives us a more low-level API to CURL without the
InputStream interface, integrated into our IOThread.