feat(transcode): add Intel VA-API (Arc) GPU backend

[Pon-node]

Adds a full-GPU Intel VA-API backend to the transcode app, alongside the existing NVENC / CPU paths. Motivated by running the transcoder on an Intel Arc A770.

What

• profiles.py — h264_vaapi / hevc_vaapi / av1_vaapi support in video_args: CQP rate control (-rc_mode CQP -qp), nv12 VA-surface pixel format, chroma clamp to 4:2:0. Selected through the existing --h264/h265/av1-encoder flags.
• engine.py — -vaapi_device for GPU encode, plus an optional HWACCEL_DECODE path (-hwaccel vaapi -hwaccel_output_format vaapi + scale_vaapi) that keeps decode + scale + encode all on the GPU. FFMPEG_BIN lets the image choose the ffmpeg build.
• Dockerfile.intel — built on jellyfin-ffmpeg7. Distro ffmpeg's scale_vaapi and av1_qsv are broken on Arc (Cannot allocate memory / Invalid FrameType); the jellyfin build ships the patched iHD + oneVPL stack that fixes both. docker-compose.intel.yml wires it up with --device /dev/dri.
• Dockerfile.gpu — NVENC CUDA base, realizing the README's existing "CUDA base" note (no functional change to the CPU/NVENC code paths).
• runner.py — advertises a non-NVIDIA GPU to orchestrator discovery via RUNNER_GPU_* env, since register_runner auto-detects NVIDIA only (Intel runners otherwise show with no gpu field).

NVENC / CPU behavior is unchanged; VA-API is purely additive and opt-in via the encoder flags + Dockerfile.intel.

Test plan

Built Dockerfile.intel and ran the batch surface on an Intel Arc A770 (renderD129):

• POST /transcode with an H264 + HEVC + AV1 ladder → all three renditions encode on the GPU, valid output (ffprobe confirms h264 / hevc / av1).
• Full-GPU AV1, 60s 1080p source, 3-rendition ladder: 1 session 17.9x realtime; 10 concurrent sessions hold 2.7x realtime each (26.9x aggregate).
• CPU: ~92% lower than the sw-decode path (decode + scale move onto the GPU), and faster wall-clock.

Notes

• The Intel path uses VA-API for every codec (one consistent scale_vaapi + *_vaapi pipeline). scale_qsv + *_qsv is an alternative that also works on jellyfin-ffmpeg, but a single VA-API pipeline is simpler and covers H264/HEVC/AV1 uniformly. On distro ffmpeg both scale_vaapi and av1_qsv fail on Arc, which is why Dockerfile.intel uses jellyfin-ffmpeg.
• No secrets or deployment-specific config included.