Fix reconnecting websockets on heartbeat timeout

[laurglia]

In Google Chrome, when network connection is interrupted, the websocket
connections break but they do not properly close. The following happens:

• The error and close events are not triggered
• Sending data with WebSocket#send works without errors even though
  the data is not actually sent
• When calling WebSocket#close, the websocket is not closed.
  That is expected, as closing a websocket includes sending a closing
  handshake which cannot be done when the underlying connection is broken.
  The closing handshake is documented in the WebSockets
  Standard.

Because of the behavior described above, Phoenix failed to reconnect
sockets with WebSocket transport. When heartbeat timed out, WebSocket#close
was used to close the connection but since the connection was broken,
sending the closing handshake failed, and the close event was not
triggered.

It seems like that after some time, after around 2--3 minutes, the
WebSocket finally closes (the close event triggers) but that is too
long period for our use case.

To demonstrate the issue I have created a toy Phoenix project which sets
up a WebSocket connection and displays Phoenix Socket logs on the page.
The project is here: https://github.com/laurglia/phoenix-reconnection-issue

To reproduce the problem:

1. Clone that repository
2. Run the Phoenix server with mix phx.server in somewhere other than
   your local network
3. Open the index route in Google Chrome
4. Interrupt the network, for example, connect to another WiFi
5. Observe that the Phoenix heartbeat fails and says that it is going to
   reconnect but in reality it does not immediately reconnect. You can use
   the "Ping" button on top of the page to also try to push a "ping" packet
   and observe that it is not sent.
6. Only after around 1-3 minutes, the WebSocket connection is
   reestablished

To make testing easier, I have set up a server which runs the toy
project. You can access the server at http://34.219.19.64:4000/. Here is
also a video of me reproducing the issue: https://youtu.be/BP3rK06p0Ww

This commit fixes the issue by "tearing down" the WebSocket connection
after heartbeat timeout instead of just attempting to close it. The
teardown functionality does not necessarily close the connection but it
will remove all references to it, allowing a new one to be created in
place of the current one.

I also needed to make some changes to the teardown functionality.
Previously, it was not necessary to remove listeners from the
connections because once a connection was closed, it was not able for it
to emit any events. However, now since the connection may actually stay
open for some time and emit a close event after a few minutes, then that
could cause problems if the connection was replaced with a new
connection. That is why we now we set this.conn.onopen,
this.conn.onerror and other listeners to no-ops when tearing down the
connection.

On my test server, I am running the updated version of Phoenix on my toy
project. You can access the updated version here: http://34.219.19.64:4001/

Here is a video of me showing that the issue has been fixed in the
updated version: https://youtu.be/3t4nckli07k

[chrismccord]

Thank you for such a detailed report and fix! I can verify that I reproduced this on Chrome, which is interesting because Safari is not affected. Give me some time to dig into the PR, but looks good at first glance. Thank you!

[laurglia]

Posting here as well for clarity but I created one more pull request to fix another, but different bug related to reconnections: #4924

[guessthepw]

@laurglia - this is amazing. I have been noticing this in my work app.

[chrismccord]

❤️❤️❤️🐥🔥