v0.10.0: Let a scene wait for one word of a spoken line.

A scene can wait for one word of a spoken line. Without a provider the
length is a speech-rate estimate, so the playground and CI still run.

Added

• Project::tts speaks through a TtsProvider. The key stays in the
  provider. A failed call falls back to the estimate and the scene still
  runs. voice("one two").start() starts the line. .at_word() waits
  until that word. Awaiting the line waits until the end. A line is
  .priority(), so it ducks a soundtrack built with .duck_under_voice().
• The recording is cached under tts/ next to the scene. Commit that
  folder to render in CI without a key. Without a provider, oxframe reads
  a committed cache and otherwise estimates the length.
• oxframe render --subtitles out.srt writes the lines. A path ending in
  .vtt writes WebVTT. --no-audio still writes them, and --range
  shifts the cues into that part of the video.