Round-trip parsing and output generation causes difference in parsed result

Found this while fuzzing.

extern crate cssparser;

use cssparser::ToCss;

fn main() {
    let input = "\\\n3:\'\\\x0c";
    println!("input:\n\t{:?}", input);

    let mut parser_input = cssparser::ParserInput::new(input);
    let mut parser = cssparser::Parser::new(&mut parser_input);
    let tokens = parser
        .next_including_whitespace_and_comments()
        .into_iter()
        .collect::<Vec<_>>();
    println!("tokens:\n\t{:?}", tokens);

    let str2 = tokens.iter().map(|t| t.to_css_string()).collect::<String>();
    println!("tokens to string:\n\t{:?}", str2);

    let mut parser_input = cssparser::ParserInput::new(&str2);
    let mut parser = cssparser::Parser::new(&mut parser_input);
    let tokens2 = parser
        .next_including_whitespace_and_comments()
        .into_iter()
        .collect::<Vec<_>>();
    println!("tokens to string to tokens:\n\t{:?}", tokens2);
}

input:
	"\\\n3:\'\\\u{c}"
tokens:
	[Delim('\\')]
tokens to string:
	"\\"
tokens to string to tokens:
	[Ident("�")]

Start by running the Rust reproducer from the issue and compare the two token lists for the shown input. Trace the token parsing and ToCss serialization paths involved; the fix is complete when parsing the generated string preserves the original parsed result and the regression is covered by a test.