borrowck is superlinear in the number of values held across an `.await`

An async fn that holds K values across a single .await spends superlinear time in mir_borrowck — doubling K costs ~6.6x. It is one dropck_outlives call on the coroutine's own type, reached from liveness::trace's dropck of boring locals. In our tree a 2000-line e2e test function costs 0.7s in that single call, and borrow checking is 64% of the crate's compile.

$ 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 hold_1000.rs
real 0.17
$ time rustc --edition 2021 --crate-type=lib --emit=metadata hold_2000.rs
real 1.15

drop_K.rs binds the same K values but reads each one before the await instead of after, so none cross a suspension point — it stays linear (0.04s and 0.07s). The types do not matter, and neither do borrows or lifetimes.

Start with the rustc-async-borrowck-repros cases hold_1000.rs, hold_2000.rs, and drop_K.rs, and reproduce their timings with the shown rustc command. Trace the reported path through mir_borrowck, liveness::trace, and dropck_outlives to identify why coroutine values held across .await scale superlinearly. Done means the hold cases scale approximately linearly without regressing the drop_K cases.