Supersedes #9316, in which @stubailo demonstrated that SockJS can be
straightforwardly eliminated if native WebSocket support is assumed.
Native WebSocket support is close to universal now (94% of all clients
according to https://caniuse.com/#feat=websockets), which means we can
remove ~30KB of legacy code from the client bundle that we send to most
web browsers.
To achieve this flexibility, the sockjs-0.3.4.js script is injected as a
standalone <script> tag in the <head> of the document, but only if the
user agent string of the HTTP request fails to convince the sockjs-shim
package that the browser has native WebSocket support. When Meteor is
running in production, a minified version of that script is used instead.
Not only is this loading strategy a bit faster for older browsers, because
the SockJS script begins loading earlier, and is independent from the
normal JS bundle, it also removes the cost of loading SockJS completely
for more recent browsers.
Though it's tempting to keep hacking the SockJS script to make it smaller,
we should avoid spending too much time on that, since it only benefits a
small portion of users whose devices are likely to have poor performance
for lots of other reasons. In other words, I think we should keep
[Amdahl's Law](https://en.wikipedia.org/wiki/Amdahl%27s_law) in mind.
edb33611e1