feat(moq-mux,moq-srt): PTS-exposing TS Export API + PCR-paced SRT egress

[kixelated]

Summary

Implements step 1 of #1839: an additive media-timing API on moq_mux::container::ts::Export, and uses it to fix the SRT egress tune-in burst that #1828 explicitly deferred.

ts::Export now yields Frame, not Bytes

• Export::next() / poll_next() return Option<moq_mux::container::Frame>. Each frame's payload holds the TS packets, stamped with the source timestamp and keyframe flag, so a transport can pace delivery on the media clock. duration is always None (the muxer never tracked it).
• PAT/PMT are buffered, then flushed with the first media timestamp. Previously the leading header went out as its own timestamp-less chunk and write_frame re-emitted PAT/PMT on the first keyframe (a startup double-PSI). Now step 3 only builds the PSI; write_frame prepends it to the first frame, so the header inherits that frame's real timestamp and timestamp is always valid (no Option). Removed the now-unused write_psi.

SRT egress paces on the media clock

• serve_request stamps each SRT payload with anchor + (timestamp - base) instead of Instant::now(). That Instant is the packet's origin time feeding SRT's TSBPD, so the receiver reconstructs the original inter-frame spacing and the tune-in keyframe burst is released at the media rate instead of bursting. One Instant::now() total (the anchor), no per-packet wall-clock stamps.

Why this is the whole SRT fix

SRT already has congestion control and a receiver jitter buffer (TSBPD). The burst wasn't a missing-pacing problem; it was that we fed Instant::now() as the packet origin time, collapsing all spacing into "simultaneous." Feeding real media time uses TSBPD as designed. The standalone PCR-pacing loop #1839 describes is still needed for the future raw UDP/RTP backends (no TSBPD there), but not for SRT.

Notes

• Export::next() returning Frame is a breaking moq-mux API change, hence targeting dev.
• duration was not removed: it's load-bearing in fmp4 export (trun sample durations) and the consumer's group-advance, and isn't needed for Export to return Frame (it just sets None, like the legacy/flv/ts-import paths).
• No cross-package sync: Rust-only muxer API (no JS equivalent), and the moq-cli CLI surface is unchanged (same TS bytes to stdout).

Test plan

• cargo test -p moq-mux ts:: (181 passed) — covers the PSI-buffering change and TS round-trips
• cargo test -p moq-srt
• cargo clippy -p moq-mux -p moq-srt -p moq-cli --all-targets (clean)
• E2E: SRT m=request tune-in no longer bursts the opening GOP (manual, ffmpeg/VLC loopback)

Part of #1839; builds on #1828.

(Written by Claude)