Bitmap subtitle burn-in (PGS/VOBSUB/DVB) ran the whole video through a
GPU->CPU->GPU roundtrip: every decoded frame was hwdownload'd to system
memory, the subtitle bitmap composited with the software overlay filter,
then hwupload'd back for the encoder. On a 1080p source that pins the
encode below realtime (~0.68x measured), so the client can never build a
buffer and rides the produced-head edge indefinitely; the same roundtrip
also intermittently crashes the QSV buffer path with SIGBUS.
Composite on the GPU instead via overlay_vaapi for QSV and VAAPI: the
decoded video never leaves its VAAPI surface and only the small,
low-frequency subtitle bitmap is uploaded. Measured ~7x realtime and
crash-free on the same file. The libass text path is unchanged (it must
stay on CPU), and NVENC/CPU keep the software overlay because overlay_cuda
is unverified on the bundled ffmpeg.
Adds QSV and NVENC bitmap burn-in tests and updates the VAAPI test to the
GPU graph.
