Faster emojify() by avoiding string.replace() entirely

[nolanlawson]

(Previously: #4046)

This speeds up emojify() by avoiding string.replace(regexp) entirely. Instead I pull in a small extra dependency I wrote, substring-trie, which implements a Trie data structure to do a prefix-based lookup that's faster than what emojione.regUnicode was doing.

Testing on a Nexus 5X in Chrome, I see the cost of emojify for switching from the "getting started" column to the local timeline go from 24.4ms to 2.3ms:

I used performance.measure() to measure the cost of building the Trie itself, and I got 18.7ms, so it's worth the tradeoff. Also substring-trie is 407 bytes minified.

Note also the extra tests, which I wrote to ensure I didn't regress anything. One detail that's important is that regUnicode returns the longest matched substring, which is also the behavior of substring-trie (and there's a test to confirm this).

[unarist]

Looks good, and this also fixes #3732.

[charlag]

Wow, that's impressive. Tries are really cool and I think it is a great use of them.