Reduce the generated code size of Timeout<T>::poll

[akonradi-signal]

Pull out the tail portion that is invariant for T into a separate function.

Motivation

This reduces the size of the built binaries for crates that use Tokio and target minimal code size.

Solution

The Rust compiler generates a separate copy of LLVM IR for each instantiation of templated functions. That means that code in a templated function that doesn't depend on the template types can be needlessly duplicated. The tail of Timeout::poll is one such function. Pulling the code that doesn't depend on the type T out into a separate function lets the compiler emit it once instead of once for each instantiation of Timeout.

[rcoh]

Seems reasonable. Do you have any data of the impact of a change like this on a real codebase? Is it actually measureable?

[akonradi-signal]

It's small. From building https://github.com/signalapp/libsignal/ with Tokio 1.45.0, the changes here only reduce the Android arm64 binary size by 3KB (out of ~6MB after stripping, though most of the code isn't async). That being said, using cargo-bloat and filtering the output for "tokio::time::timeout.*poll" shows about 10% less code text size attributed to those functions for libsignal.

[Darksonn]

This change sounds good to me. Just one nit.

[akonradi-signal]

I realized that the Result<Infallible, E> spelling was redundant and removed that in the last commit. If you'd rather keep it feel free to revert or drop that before merging.