Inline generators before state machine conversion in order to reduce branching

[pcwalton]

As of 6c51ec9, which is the fix for #76181, generators no longer inline in MIR. This results in missing the following optimization opportunity:

pub async fn f() {
    foo().await;
    g().await;
}

async fn g() {
    bar().await;
    h().await;
}

async fn h() {
    baz();
}

should become:

async fn f() {
    foo().await;
    bar().await;
    baz().await;
}

I checked a simple example and LLVM didn't seem to do the inlining, which doesn't surprise me as after the state machine transformation it's going to be quite hard to perform it.

This could be an important optimization because recreating the nested call stack when a generator is resumed can be O(n) in bad cases like this one. (This was brought up as a potential problem when comparing C++ coroutines with Rust generators.)