Video: inter-frame decode (decoded picture buffer and reference handling)

[justin13888]

rawshift-hwdec's sequence seam decodes random access points only. An access unit that references other pictures is refused with a clear error (vaapi/video.rs), so keyframe extraction, thumbnails, scrubbing and All-Intra camera modes work, but ordinary long-GOP recordings decode one frame per GOP.

Completing it is a purely internal change behind an unchanged public API — HwVideoDecoder already splits send-packet from receive-frame precisely so reordered output has somewhere to go — so it needs no second breaking release.

What it takes, per codec:

• HEVC — POC derivation (needs nuh_temporal_id_plus1, which nal_type currently discards, for prevTid0Pic); the full st_ref_pic_set derivation including inter-RPS prediction (today's parse_st_ref_pic_set only counts entries); long-term entries; RefPicSetStCurrBefore/StCurrAfter/LtCurr; reference lists; RASL discard after a CRA that starts the stream or follows a reset. parse_sps currently reads and discards sps_max_num_reorder_pics and sps_max_latency_increase_plus1; both are needed for the bumping process.
• H.264 — see the separate issue for its picture-parameter path first.
• Shared — a decoded picture buffer with surface lifetimes, the Annex C bumping process, and populating ReferenceFrames / RefPicList in the VA buffers.

Notes for whoever picks this up:

• Reorder must live in hwdec, not the caller: VAAPI is a slice-level API where the caller supplies the reference lists, so the buffer and the ordering cannot live above it.
• The DPB is a pure function over synthetic inputs and should be table-tested (POC types, MMCO, sliding window, hierarchical-B pyramids) without any GPU.
• The device test should assert bit-exactness against ffmpeg's software decode, not a variance heuristic — H.264/HEVC decode is normative, and only bit-exactness catches a wrong reference list or scaling matrix.
• Consider splitting the pure-Rust codec logic out from under cfg(hwdec_backend = "vaapi"): the test CI job runs cargo test --workspace with no features, so none of the existing parser unit tests run there today.

Follow-up to #39.