String-buffer backed TokenStream

[rylev]

This is the first steps in addressing dtolnay/quote#160

This changes TokenStreams to keep no spanned tokens as plain strings and only convert them to imp::TokenStream's when necessary.

This doesn't make a difference in the proc_macro benchmarks as that is still using older APIs that go through the proc_macro bridge, but it cuts quote's benchmark time in half. This makes my own crate's (winrt) benchmark ~35% faster, though this is not as fast as the squote experiment which did not support spanned tokens at all. The reason for this is because delimited tokens must be balanced and so we have to treat groups differently even if they only contain non-spanned inner tokens. This prevents us from simply treating groups like strings and folding them into string backed token streams.

[rylev]

@dtolnay would you be able to take a look at this?

[dtolnay]

Closing along with the quote PR since this seems to make the proc macro use case 65% slower, which is not the right tradeoff for a crate that is predominantly used in proc macros — see dtolnay/quote#162 (review).

Thanks anyway!