moq-lite-05: move immutable track props to a Track Stream (TRACK_INFO)

[kixelated]

What

Implements moq-dev/drafts#25 in rs/moq-net: move a track's immutable publisher properties out of SUBSCRIBE_OK and onto a dedicated, on-demand Track Stream. Scoped to the WIP Lite05 version; Lite01-Lite04 keep SUBSCRIBE_OK unchanged, so this is backward-compatible.

Why

Priority, Ordered, Cache, Timescale, and Compression are fixed for a track's lifetime, yet were echoed on every SUBSCRIBE_OK. Fetching them once over their own stream and caching them removes that per-response repetition (and would let group-by-group FETCHes reuse a single lookup). Keeping them immutable also avoids a relay fan-out problem: a publisher-side change would otherwise have to propagate out to every downstream subscriber.

What changed (rs/moq-net)

• New Track Stream (0x6) (lite/stream.rs, lite/track.rs): a TRACK request (broadcast path + track name) answered with a single TRACK_INFO (Priority, Ordered, Cache, Timescale, Compression), then FIN, or a reset on error / missing track. With roundtrip tests.
• SUBSCRIBE_OK slimmed (lite/subscribe.rs): the static props are gone. On Lite05 a subscription is accepted implicitly (rejection is a stream reset) and the publisher sends nothing on the subscribe stream.
• Publisher (lite/publisher.rs): recv_track/run_track_info resolve the track's .info (via a short-lived subscribe) and reply TRACK_INFO. run_subscribe no longer sends SUBSCRIBE_OK on Lite05 but still computes the same compression/timescale for serving.
• Subscriber (lite/subscriber.rs): flights TRACK and SUBSCRIBE in parallel, so the first group still arrives in one round trip. A pending TrackEntry is inserted before SUBSCRIBE, so group streams that race ahead of TRACK_INFO park on a resolved channel (buffered by QUIC flow control) instead of being dropped. The resolved (producer, compression, timescale) is reused for every group's decode and fetched once per upstream-subscription lifetime (linger), rather than re-derived per response.

Design notes / open questions

• Publisher resolves TRACK_INFO by subscribing-and-dropping. In the model, immutable props are only delivered on accept, so there's no cheap "peek info" path. At an origin this reuses the live producer instantly; at a relay it coalesces with the parallel SUBSCRIBE (linger covers the gap). A dedicated TrackConsumer::info() would make TRACK cheaper.
• Priority/Ordered are sent as 0/false. The model Track has no publisher priority/order field yet (the old SUBSCRIBE_OK echoed the subscriber's priority and hardcoded ordered=false), so nothing real is lost. Making them meaningful means adding two Track fields (+ catalog/serde + js sync).
• Not included: SUBSCRIBE_START/SUBSCRIBE_END and the SUBSCRIBE_DROP renumber. The Rust impl never resolves a group range or emits drops, so there's no functional gap, only a wire-spec one.

Cross-package sync (deferred)

Per the sync table, rs/moq-net wire changes also need js/net and doc/concept. Left out here while we gauge the approach; happy to follow up in this PR.

Test plan

• cargo test -p moq-net (350 passing) including new TRACK/TRACK_INFO roundtrip tests
• cargo clippy -p moq-net --all-targets clean
• cargo fmt (pinned nix toolchain)
• cargo check -p moq-relay -p hang -p moq-cli (downstream still builds)
• End-to-end lite session test (none exists in-crate yet; the parallel TRACK+SUBSCRIBE flow is currently only covered by unit/roundtrip + model tests)

(Written by Claude)