The pinned workers don't shut down when the tokio runtime does, so the
included test never completes.

The worker count is currently hard-coded with a fixme.

PinnedPool does not reset, so if a second runtime is started after the
first finishes, spawn_pinned will not work (the first issue about not
shutting down the workers prevents this scenario from happening atm).

This reverts commit 9d0ee8d.

This also makes spawn_pinned_blocking pretty redundant, so it is
removed.

Also added more error handling and comments. There are still two more
error handling issues:
1. `create_task` could panic, which kills the worker. Using
   `catch_unwind` limits the types you can use for `create_task` too
   much. We could use AssertUnwindSafe here, but probably shouldn't.
2. If a worker dies for some reason, the pool doesn't heal (recreate the
   worker). Healing the pool would require more mutable state and
   mutual exclusion, which might get complicated.

Co-authored-by: Alice Ryhl <aliceryhl@google.com>