Add host, Miri, and bare-metal tests plus CI

[richardanaya]

Stacked on #16 (cursor/replace-woke-with-alloc-wake-fec3), which is not merged yet. This PR is only the test suite, CI, and the small library adjustments those checks need. Retarget the base to master after #16 lands.

What the suite proves

Host tests in tests/behavior.rs (15, also run under Miri):

• tasks run to completion, including many of them
• is_done is false while work is queued or a finished task is still listed, and true after update drops it
• tasks spawned from inside a running task are picked up
• wake_by_ref and consuming wake, including from another thread, complete through update_woken then update
• update_woken does not poll tasks that were not woken
• a wake plus add_async during poll does not deadlock
• a task is not polled again after Ready
• Output types other than () (u32, String, and Cell<u32>, which is Send but not Sync)
• finished output is dropped, and the task future itself is dropped when update removes it
• a short many-wake stress run finishes quickly

no_std

nostd-test is a #![no_std] #![no_main] binary with a 16 KiB bump allocator. QEMU (lm3s6965evb, Cortex-M3, semihosting) runs it and exits 0. The harness checks, with no std linked:

• is_done starts true
• a completing task is dropped
• add_async from inside a task
• a self-wake
• eight tasks
• a u32 output

CI also cargo checks the library for wasm32-unknown-unknown and thumbv7m-none-eabi.

thumbv6m-none-eabi (Cortex-M0) is not supported. That target has no pointer-width atomics (target_has_atomic="ptr" is absent), so alloc::sync::Arc is not available and spin::Mutex cannot CAS. The same class of failure happens on main (woke fails to import alloc::sync::Arc). Not a regression. Noted in the README.

CI

.github/workflows/ci.yml on pull requests and pushes to master:

• rustfmt
• cargo clippy --all-targets -- -D warnings (and the harness for thumbv7m)
• cargo test, cargo doc --no-deps, and the four host examples
• MSRV 1.63: cargo build --lib plus the two no_std cargo checks
• QEMU run of nostd-test
• cargo miri test

The 1.63 job does not run cargo test. Current async-std releases pull crates that need a newer compiler. The library itself builds on 1.63.

Checkout uses actions/checkout@v5 (Node.js 24). v4 still targets Node.js 20, which GitHub Actions now warns about.

Library adjustments in this PR

These are here so the checks above are honest, not in #16:

• spin is pinned to 0.9.9 without lock_api. spin = "0" now resolves to 0.12, whose lock_api needs Rust 1.71, so a fresh 1.63 build could not meet the declared MSRV.
• Queues are spin::Lazy, because VecDeque::new is not const on 1.63. Same queues, built on first use.
• update_woken drains the woken queue and drops the guard before polling. spin::Mutex is not reentrant; holding it across poll deadlocked a wake from inside that poll. Wakes during poll are queued for the next update_woken.
• update drops a task in the same pass that sees it finish, and does not poll a woken task that is already done.
• The input queue is drained with while let instead of len() > 0. Stable clippy 1.99 denies clippy::len_zero under -D warnings.

examples/timer does not build: the external web crate's glob import now clashes with this crate's name, and set_inner_html is gone. Left unchanged.