Previously the ESTree tokens serializer had a special case to skip a << token which has same start as preceding < token. Instead, solve this problem at source - prevent this duplicate token from being in the Vec<Token> in the first place, by modifying the parser.

This is important because we want to move to sending tokens to JS via raw transfer, and we need the Vec<Token> to have the right number of tokens in it to start with, so we can do lazy deserialization and just get a token at a specific index. This breaks down if there are extra tokens that need to be skipped.

The logic around this in parser is quite labyrinthine, so add lengthy comments explaining it.

This change also simplifies the main loop in ESTree serializer, so has side effect of +1% on estree_tokens benchmark. Conversely, it doesn't affect the parser_tokens benchmark perf, because the path in parser which got more complex is not commonly taken.