Add wide & unicode string literal support

[ehaas]

Adds support for L"" / u"" / U"" literals.

String contents are written to p.retained_strings

Also fixes 2 unrelated bugs:

1. Max diagnostics for a char or string literal is currently capped at 4; we drop any after that. So if 4 warnings happened first, followed by an error, the error would get dropped. New behavior is to replace the last warning if adding an error fails.
2. Fix a crash when passing a non-string to an attribute that expects a string (introduced in Value: Use a start/end range instead of slice for byte values #503)

[Vexu]

Thanks for working on these! The invalid token is now almost exclusively used for incorrectly encoded bytes.

Would you mind taking a look at the remaining cases? I think some of them can simply be marked unreachable thanks to line splicing and the others could be changed to empty_char_literal and unterminated_{char,string}_literal with better error messages.

The split between CharKind and StringKind with nearly identical APIs seems a bit odd to me but if you think it's better then it's fine.

[ehaas]

I think the only reason I made them separate is to prevent string and char literals from concatenating but you're right, they can probably be combined. I'll take a look at that and the invalid token.