Make smol truly live up to its name

[ghost]

This is the final step in smol's evolution. The crate has been broken into:

• blocking, replacing smol::reader(), smol::writer(), smol::iter(), and smol::block_on().
• async-io, replacing smol::Async and smol::Timer.
• multitask, replacing smol::run().

The adapters for blocking functions have always been independent of the actual runtime, so it was natural to break those parts into blocking.

Previously, Async and Timer wouldn't work if you didn't call run(). While it was nice not to have background threads driving the runtime, this design turned out to be more pain that is worth. Now, async-io will spawn a fallback background thread driving the reactor, but executors can still drive the reactor using async_io::parking. This way we can drive the reactor and the executor on the same thread, thus getting amazing performance while still keeping things simple.

The executor in multitask is now just a state machine that users can drive on their own however they wish. All you need to do is to keep calling .tick() to run a single task at a time. This gets us much more flexibility than before. Now you can easily run two or more executors on the same thread simply by doing ex1.tick(); ex2.tick(); ex3.tick();. This opens up a world of possibilities - e.g. you can implement task priorities by creating an executor per priority level, and then "ticking" all executors on the same thread.

Smol is now just a tiny wrapper around multitask, and it works really well in small async programs and unit tests. For anything more demanding, you can trivially write your own custom smol.