Fix: Handle client connection reset for keep-alive connections

[tisba]

See #86

TCPSocket#eof? will raise Errno::ECONNRESET if the connection has been reset. This is currently not handled properly in WEBrick::HTTPServer#run. This was reported by @carlhoerberg, but incorrectly addressed in #86.

Preconditions to hit this issue:

• Client has to request with Connection: keep-alive
• WEBrick has to respect that (e.g. errors / 404s do close the connection)
• Client resets the connection after the first request

This will leave the server believe that the connection is still good and fail in WEBrick::HTTPServer#run when testing TCPSocket#eof? in the next iteration of #run.

The fix is to rescue Errno::ECONNRESET when checking for #eof?.

Testing this is a little tricky, here is how I approached it:

• start a test server, using WEBrick::HTTPServer
• setup a success response on /
• do a single request with Connection: keep-alive
• after reading the response body, fish the TCPSocket object out of the Net::HTTP instance and configure it to linger on close and forcefully reset the connection, dropping all remaining data. This will send a RST to the server
• to avoid TestWEBrick shutting down, before we hit the issue, we need to wait a little bit

[tisba]

Can somebody kick of the test workflow? @olleolleolle maybe? :)

[jeremyevans]

Thank you for working on this! The approach you used for fixing the issue seems best to me.