refactor(test): execute all #[rustup_macros::unit_test]s within a tokio context

[rami3l]

Split from #3803.

With this PR, all tests under the #[rustup_macros::unit_test] attribute now execute within a tokio runtime, eliminating the need of static TRACER and TRACE_RUNTIME variables. Also, the setup of tracing subscribers in both bin and test targets has also been unified, to prepare for the addition of another stderr-oriented subscriber in #3803.

Rationale

Currently, the tracing subscriber registration will fail if a tokio runtime is not present. To mitigate this problem, static TRACER and TRACE_RUNTIME variables are added as an ad-hoc fallback if we're running a non-tokio unit test.

rustup/src/test.rs

Lines 230 to 237 in 3ba08da

	/// A tokio runtime for the sync tests, permitting the use of tracing. This is
	/// never shutdown, instead it is just dropped at end of process.
	#[cfg(feature = "otel")]
	static TRACE_RUNTIME: Lazy<tokio::runtime::Runtime> =
	Lazy::new(|| tokio::runtime::Runtime::new().unwrap());
	/// A tracer for the tests.
	#[cfg(feature = "otel")]
	static TRACER: Lazy<opentelemetry_sdk::trace::Tracer> = Lazy::new(|| {

However, the fact that the subscriber state is global in tests has made some unnecessary complications especially WRT adding other tracing subscribers, such as the stderr-oriented tracing subscriber proposed in #3803 (comment), which is bound to the local currentprocess:process() context (see #3803 (comment)).

I decided to go with local contexts all the way, and it did work. However, the main downside to this approach is that #2367 has to be reverted, since run_inprocess() calls with_runtime(), and for the latter to correctly alter static PROCESS in-process, a tokio runtime builder is required, which in this case will cause a "runtime-in-runtime" error once the builder is run.

Fortunately, no obvious slowdown has been observed WRT our CI so far. Also, as now main() seems to be the only caller to with_runtime(), this gets us closer to using #[tokio::main].