• Update voices.json: 6 new voices, 36 updated. The Japanese voice is now
  ja_JP-hi_fi_captain-medium; the old ja_JA-hi_fi_captain-medium name is
  kept as an alias
• Add --upload-dir, where the web UI saves uploaded Piper voices (default:
  --download-dir, as before). Uploads and catalog downloads used to share
  --download-dir, so a backup could not include the irreplaceable uploads
  without also carrying every re-downloadable catalog voice (20-110 MB each).
  With separate directories, the download directory can be left out of backups
  whole. An upload directory that is not a --data-dir is searched before
  all of them, so uploads are always found and a custom voice wins over a
  catalog voice of the same name
• Answer a plain synthesize sent on the same connection after a text stream.
  The streaming flag set by synthesize-start was never cleared, so every later
  synthesize was skipped as a compatibility copy and the client waited for an
  audio-stop that never came