fix(irodori-tts): mask empty captions and guard the tail trim

Two divergences from the reference implementation, both affecting
caption-conditioned checkpoints (v2 VoiceDesign, v3 VoiceDesign, v4):

- An absent caption still tokenizes to a BOS token, and its mask was
  left set, so the DiT and the duration predictor saw a one-token
  caption instead of no caption at all. Upstream zeroes the caption mask
  when the caption text is empty. For v4 voice cloning without a caption
  this shortened the predicted duration by ~4% (79 vs 82 frames) and cut
  the tail of the utterance.
- The trailing-silence trim was applied unconditionally, so a silence
  point of 0 would have produced zero-length audio. Upstream only
  applies the trim when the flattening point is positive.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>