tracing: significantly improve code generation at trace points

[jstarks]

ValueSets contain both a FieldSet reference and a slice of (&Field, Option<&dyn Value>) pairs. In cases where ValueSets are generated via documented interfaces (specifically, tracing::event! and other macros), the Field references are redundant, because the ValueSet contains a value slot for every field (either a value or None), in the correct order.

As a result, the code generated by the macros is terrible--it must put a Field on the stack for each field--that's 32 bytes per field! This is a lot of work for apparently no purpose at runtime, and it can't be moved into a read-only data section since it's intermixed with dynamic data.

Fix this by adding a variant of ValueSet that skips the Field references, knowing that it represents the full set of fields. Keep the old kind of ValueSet, too--it's still needed by Span::record, by old versions of crates, and potentially by third-party crates using undocumented methods such as FieldSet::value_set.

In some adhoc tests on x86_64 Linux, this reduces the code size as follows:

• One-field event: 258 bytes to 189 bytes, 25% reduction.
• Five-field event: 638 bytes to 276 bytes, 57% reduction.
• In a larger project with lots of events, ~5% reduction in .text section.

[jstarks]

Test and clippy failures are pre-existing. But the test failure is a little alarming since it's in the area that I'm changing. Looking into it.

[jstarks]

I see the test bug. Someone regressed the test last year while cleaning up warnings. I will open a separate PR to fix this.

[jstarks]

#3399