## Motivation
I was just trying to use tokio-trace for a greenfield project, but I was frustrated to discover that I couldn't really use it easily.
I was using the [`runtime`](https://docs.rs/runtime/0.3.0-alpha.4/runtime/) crate, which transparently spawns a thread pool executor for futures. In that thread pool, there's no way to set a tokio-trace subscriber for the duration of each thread, since you don't control the thread initialization. You *might* be able to wrap every future you spawn with a subscriber call, but that's a lot of work.
I was also confused because the documentation said that setting a subscriber in the main thread would use that subscriber for the rest of the program. That isn't the case, though -- the subscriber will be used only on the main thread, and not on worker threads, etc.
## Solution
I added a function `set_global_default`, which works similarly to the `log` crate:
```rust
tokio_trace::subscriber::set_global_default(FooSubscriber::new());
```
The global subscriber (actually a global `Dispatch`) is a `static mut` protected by an atomic; implementation is copied from the `log` crate. It is used as a fallback if a thread has no `Dispatch` currently set. This is extremely simple to use, and doesn't break any existing functionality.
Performance-wise, thread-local `Dispatch` lookup goes from ~4.5ns to ~5ns, according to the benchmarks. So, barely any runtime overhead. (Presumably there's a little compile-time overhead but idk how to measure that.) Since the atomic guard is only ever written once, it will be shared among a CPU's cores and read very cheaply.
I added some docs to partially address #1151. I also switched the tokio-trace benchmarks to criterion because the nightly benchmarks weren't compiling (missing `dyn` flags?)
