borrowck is superlinear in the number of `.await`s in one function

An async fn with many awaits spends superlinear time in mir_borrowck — doubling the awaits costs ~9x. It is the EverInitializedPlaces fixpoint: it visits each basic block about N/2 times where N is the await count, while every other dataflow analysis over the same body settles in one or two. In our tree one 3539-line e2e test function with 679 awaits cost 8.4s of mir_borrowck on its own.

$ git clone https://github.com/ashi009/rustc-async-borrowck-repros && cd rustc-async-borrowck-repros
$ rustc -V
rustc 1.97.1 (8bab26f4f 2026-07-14)
$ time rustc --edition 2021 --crate-type=lib --emit=metadata awaits_800.rs
real 2.27
$ time rustc --edition 2021 --crate-type=lib --emit=metadata awaits_1600.rs
real 21.67

Each .await leaves a loop in the MIR borrowck sees, and loops_N.rs — a plain fn with N sequential loop { … break } blocks and no async — shows the same per-block visit count, so the apparent trigger is loop count rather than the coroutine. Same timings on 1.98.0-beta.2 and 1.94.1.

Start with mir_borrowck and the EverInitializedPlaces fixpoint, then reproduce the scaling with awaits_800.rs and awaits_1600.rs from the linked repository. Compare the per-block visits against loops_N.rs, and consider the issue resolved when borrow checking no longer scales superlinearly with the number of awaits or equivalent loops.