Make the fast path take no locks

Ideally we'd like event recording to not even take the sharded read lock on the fast path. This _should_ be possible by keeping the TLS in some kind of atomic linked list instead of as a map inside the critical state. There are two main challenges with this approach:

- How do we deal with `span::Id` re-use and invalidation? If the writers kept a cache of the state for spans in their TLS, they wouldn't know when that state is out-of-date if a `span::Id` is dropped and then re-created rapidly. We could conceivably switch away from `Slab` and just use a monotonically increasing `u64` as the `span::Id` instead, though it would slow down lookups as they'd now need to navigate a map of some kind.
- It precludes one thread (most likely the one that'll read the histograms) from dropping all the recorders to force synchronization of the histograms. Maybe we're okay forgoing that feature?