Reduce the cost of initializing a fresh OnceCell #398

A fresh OnceCell<usize> with an immediately ready initializer took 278 ms for six million initializations. The gap narrowed and the fastest peer changed when the initializer was already running and four other callers were waiting.

Apple M4 Max, macOS arm64, rustc 1.101.0-nightly. Each cell gives rank, elapsed time relative to Asyncband and the whole-batch median over 10 interleaved runs. Lower is faster. Supplemental measurements

get_or_try_init checks the value, acquires a semaphore permit, checks again and publishes the result. An allocation counter measured one 64-byte allocation per cold initialization on this macOS host and zero for Tokio and async-once-cell. The first lock of a fresh std::sync::Mutex produced the same allocation. In a separate controlled run, changing only the internal lock implementation to parking_lot reduced Asyncband from 284 to 115 ms, with async-once-cell at 61 ms. The cost of already initialized reads is a separate measurement.

First contribution

Reproduce the cold and four-waiter rows and add an initialized-read control. Count allocations separately from timing. A focused patch can then target work done before any second caller arrives. Preserve publication of one value, retry after a cancelled or failed initializer, and progress for waiting callers.

To run the supplemental peer_checks command below, save and run peer-checks.rb from the linked comment in a disposable checkout of 6102089. The benchmark and library sources match 36ba6f2. Generator and instructions

cargo x bench --bench primitives -- once::get_or_init::initialize_cell --repeat 10 --raw
cargo x bench --bench peer_checks -- once-cell/initialize_cell --repeat 10 --raw

Performance tracking issue #396