Improve performance of inactive tracing spans

[TheJokr]

Improve performance of log/tracing harness access

• Switch to std::sync::{OnceLock, LazyLock} for better storage
  efficiency and std's futex-based backend. These are available since
  Rust 1.80.
• Avoid unnecessary LazyLock initialization of NOOP_HARNESS if a
  proper harness is configured. Previously, .unwrap_or(&NOOP_HARNESS)
  would cause the Deref impl to be called even if the reference was
  never used. (Due to Deref's side effects.)
• Wrap harnesses in crossbeam's CachePadded to avoid false sharing
  with other global variables. This is important to ensure OnceLock's
  initialization check is not slowed down by dirty cache lines.

Remove globally-shared INACTIVE_SPAN RwLock

This RwLock was a heavy contention point in applications that generate a
lot of spans, as every span operation must lock and unlock the RwLock.
Each lock/unlock is an atomic RMW and thus dirties the cache line for
all other cores.

We can avoid the global singleton INACTIVE_SPAN entirely with some
changes to our internal SharedSpanHandle API.

Remove internal once_cell usages

There are two spots left where we can't remove once-cell yet:

• In memory_profiler.rs, we use OnceCell::get_or_try_init. This
  function is not stabilized in std yet.
• The foundations::security::allow_list macro generates OnceCell
  instances in user code. Changing this requires a major bump.