connections graceful shutdown

[icing]

Implement graceful shutdown of connections in libcurl:

• when a connection is ready to be discarded, run a non-blocking shutdown on all connection filters. If that returns errors, discard the connection. If that sucessfully shuts down all filters, discard as well. Otherwise, add it to a "shutdown" list in the transfers connection cache.
• limit the shutdown list to the same maximum length as the connection cache's live limit. When shutdown down a new connection at the limit, the oldest one is discarded.
• In multi_wait() and multi_waitfds(), collect all connection caches involved (each transfer might carry its own) into a temporary list. Let each connection cache on the list contribute sockets and POLLIN/OUT events it's connections are waiting for.
• in multi_perform() collect the connection caches the same way and let them peform their maintenance. This will make another non-blocking attempt to shutdown all connections on its shutdown list.
• when a connection cache is destroyed, discard all connections on the shutdown list. In debug builds, with the envionment variable CURL_GRACEFUL_SHUTDOWN, run a blocking shutdown (using poll) and the default 2 second timeout. This is for testing (see below) and might be exposed as an option to new libcurl API function.

TODO: multi_socket() does not know how to handle sockets from connection caches.

Details

• similar to connected add a shutdown bit at the connection filter instance to remember it has already shutdown and need not be invoked again when we do repeated, non-blocking attempts to shut down the complete filter chain.
• conncache structs have a tmp_list member that allows collecting caches onto a temporary list inside multi_wait and multi_perform handling. The list is always cleared before calls return.
• replace Curl_discard(data, conn) with `Curl_conncache_shutdown_conn(data, conn,...) to handle keeping connections for non-blocking shutdown handling.
• When CURL_DEBUG is in the environment and curl is a debug build, make the conncache closing handle verbose.
• fix openssl shutdown to be way more careful about still receiving data after its side has shut down. HTTP/2 connections may have a server-side GOAWAY in transit. We need to cope with that without error.

Testing

• add test cases using tcpdump to check if a curl run caused TCP RST packets. These checks are used in test_02_30, test_30_06 and test_31_06.

[icing]

Closing to make an updated, squashed one.