Sticking to the ASCII character set (plain text, 7-bit characters) is preferred in GNU source code comments, text documents, and other contexts, unless there is good reason to do something else because of the application domain. For example, if source code deals with the French Revolutionary calendar, it is OK if its literal strings contain accented characters in month names like “Floréal”. Also, it is OK (but not required) to use non-ASCII characters to represent proper names of contributors in change logs (see Change Logs).
If you need to use non-ASCII characters, you should normally stick with one encoding, certainly within a single file. UTF-8 is likely to be the best choice.