Underscore (ASCII 95) should be encoded to maintain reversibility

Nobody has claimed this yet.

It appears that a literal underscore (ASCII 95) `_` is left as-is, meaning that a space (ASCII 32) and underscore are indistinguishable in the Q-encoded result.

Example:
```
var q = require("q-encoding")
var utf8 = require('utf8');
q.encode(utf8.encode("❤️ Hello _it's _me"));
```

Expected result:
```
"=E2=9D=A4=EF=B8=8F_Hello_=5Fit=27s_=5Fme"
```

Observed result:
```
"=E2=9D=A4=EF=B8=8F_Hello__it=27s__me"
```

Further example:
```
utf8.decode(q.decode(
q.encode(utf8.encode("❤️ Hello _it's _me"))
));
```

Result:
```
"❤️ Hello it's me"
```

RFC 2047 4.2(3) notes
> (3) 8-bit values which correspond to printable ASCII characters other
than "=", "?", and "_" (underscore), MAY be represented as those
characters.

So `_` is a special purpose character in the encoding (representing a space), so a literal `_` needs to be encoded *as if it were* non-ASCII.

Start at the q.encode and q.decode entry points and reproduce the underscore examples from the issue. Verify that literal underscores are encoded distinctly from spaces, and that encoding followed by decoding preserves the original string and matches the expected output.

Written by the indexing model from the issue text.