String-buffer backed TokenStream

[rylev]

This addresses #160 along with dtolnay/proc-macro2#239.

The bench mark runs more than twice as fast, and my own internal benchmarking shows improvements, but there are limitations to potential performance gains discussed in the proc-macro2 pull request.

[jplatte]

I'd be very interested in trying this out via [patch] but unfortunately sth. in my dependency tree has a minimum version requirement for quote of 1.0.8. Can you rebase this (and dtolnay/proc-macro2#239)? Might also increase the likelyhood of having it merged. (though general lack of maintainer attention is clearly the larger problem)

[dtolnay]

(Rebased: 0c1f124 + dtolnay/proc-macro2@9a035f2.)

According to the quote! benchmark in https://github.com/dtolnay/quote/tree/8f0d3c19eb1caba32965b551e99c00b68568cec5/benches, it looks like this pair of changes ends up being 65% slower in the proc macro case than current master. I don't think this is going to be the right tradeoff for a crate that is predominantly used in proc macros.

Thanks anyway for putting this together! I still consider it very likely possible to do string buffering in a way that makes quote faster in both macro and non-macro context, because calls over the compiler's proc macro bridge are moderately expensive and can be reduced by batching in a string buffer that's parsed all in one call over the bridge.

I'll close the PR to preserve this version of things but am open to another PR that does not regress the proc macro use case.