`commonmark.js` should first evaluate character references, before interpreting “words” of the info string

Here are some examples:

```	js
```
``` js
```
``` js
```
```js	
```
```js 
```
```js 
```
```js	a
```
```js a
```
```js a
```

While the spec is somewhat vague, I believe that the encoded whitespace in the first two examples, should behave like spaces, which is what happens with all the other examples, and that that is the intent of the spec.
Github does treat all these examples as js, and adds their class to it. CommonMark.js currently yields:

<pre><code></code></pre>
<pre><code></code></pre>
<pre><code class="language-js"></code></pre>
<pre><code class="language-js"></code></pre>
<pre><code class="language-js"></code></pre>
<pre><code class="language-js"></code></pre>
<pre><code class="language-js"></code></pre>
<pre><code class="language-js"></code></pre>
<pre><code class="language-js"></code></pre>

I expect <pre><code class="language-js"></code></pre> for everything (like GH).

The issue provides fenced-code examples for commonmark.js and the expected HTML output; start by reproducing those examples with the parser. Trace how the info string is interpreted and character references are evaluated, then add coverage showing that each encoded whitespace case produces the expected language-js class.