Avoid creating a closure for each log message

[c42f]

Closures aren't strictly required for delayed log message creation, so we can remove them to avoid #15276 and reduce the load on the compiler.

This should fix #25909 (CC @maleadt).

[Edit: changed the description for clarity]

[maleadt]

Ah, too bad, delayed message was a valuable feature for me. Maybe a secondary API (@if_debug), or keyword option to eg. @debug, could offer similar functionality?

[JeffBezanson]

Would it make sense to create a closure only if the message requires computation, i.e. isn't a constant string?

[c42f]

There's a misunderstanding here (my fault for being unclear): this change doesen't modify the public logging API and message computation is still deferred until shouldlog(current_logger(), some_static_info) returns true.

The only thing which changes is an implementation detail:

• Create a message generation closure and pass it to an internal function which later executes it, vs
• Just insert the try/catch inline in the caller, along with the message generation code.

[c42f]

CI errors seem unrelated (I restarted the OSX build which had a weird failure in the spawn tests, probably wasn't worth it though).

If there's no objections to the newly cleaned up explanation I'll merge this in the next day or so.