withoutboats has explored these trade-offs on their blog, e.g.
https://without.boats/blog/poll-drop/ and as part of io-uring investigation. io-uring is the main use-case for async drop, but even that has been mostly solved without language changes: there is tokio-uring now.
Apart from requiring a safe zero-copy io-uring abstraction to use a buffer pool instead of "bare" references, Rust's polling+cancellation model is fantastic. It's incredibly convenient to be able to externally time out and abort any Future, without needing to pass contexts/tokens and/or needing every I/O operation to have an explicit timeout option.
It's annoying when conscious small trade-offs are framed as fatal flaws (like the title of TFA). In the case of the cancellation model the "doesn't work" is that a safe abstraction is not entirely zero-cost. But when you use async in JS, C#, or Golang, you already have unavoidable heap allocations and memcpys (and another set of difficulties with cancellation). Rust's worst-case flaw is comparable to a normal mode of operation in a few other async languages. The scale of the issue is overblown, and its benefits are overlooked.