[loose parser] Fix interpretation of f."

Before this the ast produced by parse_dammit crashed in the following
code, as Uglify correctly noticed that f."" is invalid.

    sample = 'f."';

    loose = require('acorn/acorn_loose');
    uglify = require('uglify-js');

    out = new uglify.OutputStream();
    ast = loose.parse_dammit(sample);
    ast = uglify.AST_Node.from_mozilla_ast(ast);
    ast.print(out);
    // TypeError: Cannot call method 'toString' of undefined
    // member_exp.computed = false && member_exp.property == ""

    console.log(out.toString());

After this the round-tripped AST looks like: `t.✖;"";`, which is
consistent with how `foo.{` is parsed.

I also considered making it parse as t[""], but as this only turns up in
the wild when people try to use multiline strings, I felt it was better
to be obviously wrong.

acorn_loose.js

@@ -606,7 +606,7 @@
         if (curLineStart != line && curIndent <= startIndent && tokenStartsLine())
           node.property = dummyIdent();
         else
-          node.property = parsePropertyName() || dummyIdent();
+          node.property = parsePropertyAccessor() || dummyIdent();
         node.computed = false;
         base = finishNode(node, "MemberExpression");
       } else if (token.type == tt.bracketL) {
@@ -735,6 +735,10 @@
     if (token.type === tt.name || token.type.keyword) return parseIdent();
   }
 
+  function parsePropertyAccessor() {
+    if (token.type === tt.name || token.type.keyword) return parseIdent();
+  }
+
   function parseIdent() {
     var node = startNode();
     node.name = token.type === tt.name ? token.value : token.type.keyword;