tracing: move macro callsite impls out of macro expansion

[hawkw]

Motivation

Currently, every tracing macro generates a new implementation of the
Callsite trait for a zero-sized struct created for that particular
callsite. This callsite accesses several statics defined in the macro
expansion.

This means that each tracing macro expands to a lot of code — check
out the cargo expand output:

eliza on butterfly in tracing/examples on  master [$?] is v0.0.0 via ⚙️ v1.44.0
:; cargo expand --example all-levels | wc -l
    Checking tracing-examples v0.0.0 (/home/eliza/code/tracing/examples)
    Finished check [unoptimized + debuginfo] target(s) in 0.20s

463

More code in the macro expansion means more code in the function
invoking the macro, which may make that function harder for rustc to
optimize. This effects the performance of other code in the function,
not the tracing code, so this isn't necessarily visible in tracing's
microbenchmarks, which only contain tracing code.

In rustc itself, there is a small but noticeable performance impact
from switching from log to tracing even after making changes that
should make the filtering overhead equivalent:
rust-lang/rust#74726 (comment).
This appears to be due to more complex generated code impacting
optimizer behavior.

Solution

This branch moves the callsite generated by each macro out of the macro
expansion and into a single MacroCallsite private API type in the
__macro_support module. Instead of creating a zero-sized Callsite
static and multiple statics for the Metadata, the Once cell for
registration, and the Interest atomic, these are all now fields on the
Callsite struct. This shouldn't result in any real change, but makes
the implementation simpler. All the hot filtering functions on
MacroCallsite are #[inline(always)], so we shouldn't be adding stack
frames to code that was previously generated in the macro expansion.

After making this change, the expanded output is about half as long
as it was before:

eliza on butterfly in tracing/examples on  eliza/smaller-macros [$?] is v0.0.0 via ⚙️ v1.44.0 
:; cargo expand --example all-levels | wc -l
    Checking tracing-examples v0.0.0 (/home/eliza/code/tracing/examples)
    Finished check [unoptimized + debuginfo] target(s) in 0.75s

233

This change appears to fix most of the remaining rustc performance
regressions:
rust-lang/rust#74726 (comment)

Additionally, it has some other side benefits. I imagine it probably
improves compile times a bit for crates using tracing (although I
haven't tested this), since the compiler is generating fewer callsite
implementations. Reducing the number of branches in the macro expansion
probably helps make the pesky cognitive_complexity Clippy lint show up
less often, and improves maintainability for the macros as well.

[hawkw]

We could factor more code out of the macros by encapsulating the ValueSet construction in a closure that's passed to some function. However, adding a closure to every callsite probably makes build times worse, so I'd prefer to avoid this unless necessary.

[davidbarsky]

LGTM, but just asking some corrections on the doc (private) comments.

[yaahc]

😎