refactor(net)!: drop moq-net's direct tokio dependency

[kixelated]

Summary

Continues the tokio removal that started with the caller-driven (Session, Driver) refactor (#2302). This pass removes every direct tokio use from moq-net, so the crate's only remaining runtime tie is timers via web_async::time (tokio-backed on native, wasmtimer on wasm). On the wasm target, moq-net's dependency tree now contains no tokio at all.

• All 28 tokio::select! sites are now explicit kio::wait poll closures that call the model's poll_* methods directly (poll_requested_track, poll_next, poll_unused, poll_next_frame, poll_read_chunk, ...), pinning a future via waiter.poll_future only for compound work or transport-trait methods that have no poll form (Session::closed, Reader::decode, timers). New poll_closed methods on track::Producer, group::Producer, and broadcast::Dynamic fill the gaps, with the async closed() now wrapping the poll per the usual plumbing. The one shared helper that remains is util::err_only, which parks a fallible future on Ok (replacing the Err(e) = fut select pattern-guard arms in the session drivers).
• tokio::sync::watch channels became kio::Producer/Consumer<u8> with a last-observed value on the reader (lite::priority, the publisher's track-priority fan-out). send_if_modified maps to compare-before-write so unchanged values still don't wake.
• tokio::sync::Notify in the ietf Control (MAX_REQUEST_ID flow control) and the watch-based PeerSetup slot became kio::Shared state.
• The ietf control-stream adapter's mpsc channels and async Mutex became a small Queue over kio::Shared (the reverse-queue pattern from its docs): virtual streams push, the writer task drains, and a single-owner QueueWriter's drop FINs the matching recv stream exactly like the old sender-drop. accept_bi callers park on the shared queue directly, which eliminates the async mutex around the receiver, and the adapter owns its control queue instead of having the session thread a channel through.
• tokio::pin! became std::pin::pin!; tokio moves to dev-dependencies (tests still use #[tokio::test] and paused time), and the unused io-util feature is gone.
• Biasing note: tokio::select! randomizes arm order; the poll closures check in declaration order. The order matches the previous biased; sites and puts closure/error watchdogs first elsewhere, which is deterministic and at worst equivalent.

Public API changes

• Breaking: removed impl From<tokio::time::Instant> for moq_net::Timestamp. It existed only because web_async::time::Instant is tokio's Instant on native; From<std::time::Instant> remains. Hence the dev target.
• Additive: poll_closed on track::Producer, group::Producer, and broadcast::Dynamic (public poll counterparts of the existing closed()).
• Everything else touched is private or pub(crate) (PeerSetup, Control, PriorityHandle, the adapter internals, util).

Cross-package sync

No wire change (no draft update) and no js/net mirror needed: the JS side has no tokio equivalent, and no message, field, or negotiation changed. doc/concept doesn't document the runtime requirement; the crate-level Async docs in lib.rs and the rs/CLAUDE.md poll-plumbing note were updated instead.

Test plan

• just rs check (fmt, clippy -D warnings, all tests) is green; cargo test -p moq-net passes all 492 unit tests, including the priority-queue notification and virtual-stream tests that exercise the kio replacements across tasks.
• RUSTDOCFLAGS="-D warnings" cargo doc -p moq-net is clean.
• cargo check --workspace --all-targets confirms no dependent crate used the removed From impl.
• The wasm target (cargo check -p moq-wasm --target wasm32-unknown-unknown) fails identically on the base commit with a pre-existing local-toolchain error, so it gates on CI's nix build as usual.

🤖 Generated with Claude Code

[sourcery-ai]

Sorry @kixelated, you have reached your weekly rate limit of 500000 diff characters.

Please try again later or upgrade to continue using Sourcery