Make entrypoints more efficient part 2

[udoprog]

This is a different stab at #4513.

We use syn to parse the signature into our own ItemFn, maintaining most of the existing API but avoiding parsing (and reparsing) all statements except the parts we are interested in.

For signature parsing, we rely almost entirely on syn. The only notable trouble is the parsing of inner attributes, for which I believe I've matched syn's behavior: The inner attributes are expanded as outer attributes instead. In theory we could expand them as inner attributes to the newly generated function but that doesn't seem to be necessary and I did it like this to match the previous expansion more precisely.

As a basic test, I modified the tinyhttp example and put everything in the main function, these were the timing differences I got for the macro:

Before:

[tokio-macros/src/entry.rs:466] std::time::Instant::now().duration_since(start) = 13.318901ms

After:

[tokio-macros/src/entry.rs:447] std::time::Instant::now().duration_since(start) = 1.075437ms

Note: We're still not down to microseconds like the other parsing method, but now at least we don't have to worry about signatures and how to remove syn as a direct dependency without msrv stuff breaking 😄

As before, this is most notable for me when I was working on large entrypoints in rust-analyzer, for which large kinds of input latency caused diagnostics to be noticeably delayed. I've since refactored all of my entrpoints to quickly call another function to avoid this, but this should still be valuable for the future.

[taiki-e]

LGTM. This is more robust and much easy to maintain (and understand) than #4513.

[taiki-e]

Thanks!