viser v0.5.1

Correctness fixes backported onto the 0.5 line. (The NVENC capped-CRF fix and the
per-shot vmaf_model/allow_hdr option from 0.7.1 are not included — the former
targets code introduced after 0.5, the latter is an enhancement rather than a fix.)

Fixed

• Shot detection — scdet parsing now keys off the lavfi.scd.time cut flag
  instead of treating every per-frame score as a boundary. Previously every frame
  was marked a boundary and the minimum-duration merge collapsed them back into a
  single shot, so a 60s clip with 10 cuts reported just 1. detect() now also fails
  loudly on ffmpeg errors and skips boundaries at/after the total duration.
• Convex hull — non-finite (NaN/inf) bitrate/VMAF points are filtered before hull
  construction, avoiding a panic.
• extract() — validates that start is non-negative and duration is positive and
  finite; concat list paths now escape backslashes as well as single quotes.
• Probe (MediaInfo path) — the 8-bit chroma fallback emits a valid yuv420p
  pixel format instead of the invalid yuv420p8le.
• Quality measurement — frame-extraction and XPSNR failures include ffmpeg stderr,
  and a frame-count mismatch is warned. Weighted PSNR (6·Y + U + V) / 8 is only used
  when both chroma planes are present, otherwise it falls back to luma.
• Per-segment — guards against zero total duration so average bitrate/VMAF are
  0.0 rather than NaN.
• CLI loudness report — includes true-peak (Peak:) and gating threshold lines;
  the over-broad starts_with("L") filter that dropped them while printing header
  noise is fixed.
• Comparison player — frame timestamps derive from the source's real frame rate
  (probed) instead of a hardcoded 24 fps, fixing dip seek positions on 25/30/50/60 fps
  content.

Added

• per-segment analyze --segment-duration — segment length is configurable
  (default 1s) instead of hardcoded.