Improve debug mode perf of proc_macro wrapper

[dtolnay]

I dropped the following code (below; extracted from https://github.com/rylev/winrt-profile) into a dummy proc macro crate to measure the impact on winrt's code generation.

Basically it seems there are two things it's sensitive to:

• Reducing the number of calls over the proc macro bridge (roughly: number of times anything in real proc_macro gets called) can significantly improve performance.
• Reducing various overheads in the proc-macro2 wrapper layer can significantly improve performance.

This PR includes two commits, one of each type. There are certainly more of each left to do. So far in total this improves expansion time by 10%.

---

use proc_macro2::TokenStream;
use std::fs;
use std::time::Instant;
use winrt_gen::{NamespaceTypes, TypeLimit, TypeLimits, TypeReader, TypeStage, WinmdFile};

let mut paths = Vec::new();
for f in fs::read_dir("winmds").unwrap() {
    let path = f.unwrap().path();
    paths.push(WinmdFile::new(path));
}
let begin = Instant::now();
let reader = TypeReader::new(paths);
let mut limits = TypeLimits::new(&reader);
let types = NamespaceTypes {
    namespace: "windows.ui.xaml".into(),
    limit: TypeLimit::All,
};
limits.insert(types).unwrap();
let stage = TypeStage::from_limits(&reader, &limits);
let tree = stage.into_tree();
let out: TokenStream = tree.to_tokens().collect();
println!("ELAPSED: {:?}", begin.elapsed());

Ref dtolnay/quote#160

[rylev]

@dtolnay thanks for this! I'm seeing similarish performance increases on my machine. One area where I think we're crossing proc_macro bridge is when parsing strings as TokenStreams. Most parses take less than ~30microseconds to complete but we do this a lot. Is this an area you think we can improve on?