Load procedural macros asynchronously

[matklad]

Investigation in #8623 showed that switching the crate graph is slow, and procedural macro loading is a significant contributor. In particular,

https://github.com/matklad/rust-analyzer/blob/8d54fd105cbeb159c39fbc803ed45d450f0fe000/crates/proc_macro_api/src/lib.rs#L79-L82

is called on the main loop for each proc macro, and it's slow, because it reads proc-macro file to get a version, and then waits for the proc-macro server process to dlopen the file and report back.

We should move all this work off the main loop.

Plan:

• split proc-macros out of crate graph and into a separate salsa query. This is useful in general for "reload proc macros" functionality.
• make to_crate_graph return a crate graph and proc macros separately
• switch to the new salsa representation, while keeping the current synchronous loading
• (the tricky bit) add the main_loop Task to handle loading of proc macros asynchronously