Fix incorrect location data in OUTDENT nodes

In f609036bee, a line was changed from
`if length > 0 then (length - 1) else 0` to `Math.max 0, length - 1`. However,
in some cases, the `length` variable can be `undefined`. The previous code would
correctly compute `lastCharacter` as 0, but the new code would compute it as
`NaN`. This would cause trouble later on: the end location would just be the end
of the current chunk, which would be incorrect.

Here's a specific case where the parser was behaving incorrectly:
```
a {
  b: ->
    return c d,
      if e
        f
}
g
```

The OUTDENT tokens after the `f` had an undefined length, so the `NaN` made it
so the end location was at the end of the file. That meant that various nodes in
the AST, like the `return` node, would incorrectly have an end location at the
end of the file.

To fix, I just reverted the change to that particular line.

lib/coffee-script/lexer.js

@@ -768,7 +768,7 @@
       }
       locationData = {};
       ref2 = this.getLineAndColumnFromChunk(offsetInChunk), locationData.first_line = ref2[0], locationData.first_column = ref2[1];
-      lastCharacter = Math.max(0, length - 1);
+      lastCharacter = length > 0 ? length - 1 : 0;
       ref3 = this.getLineAndColumnFromChunk(offsetInChunk + lastCharacter), locationData.last_line = ref3[0], locationData.last_column = ref3[1];
       token = [tag, value, locationData];
       return token;

src/lexer.coffee

@@ -658,7 +658,7 @@ exports.Lexer = class Lexer
 
     # Use length - 1 for the final offset - we're supplying the last_line and the last_column,
     # so if last_column == first_column, then we're looking at a character of length 1.
-    lastCharacter = Math.max 0, length - 1
+    lastCharacter = if length > 0 then (length - 1) else 0
     [locationData.last_line, locationData.last_column] =
       @getLineAndColumnFromChunk offsetInChunk + lastCharacter