[KakaruHayate]

What

Two small consistency fixes for the word-level duration/word-division inputs used by DiffSinger variance/pitch models:

1. word_dur uses cumulative rounding. DiffSingerBasePhonemizer used loor(end_ms / frameMs) - floor(start_ms / frameMs) for each word interval. Training builds ph_dur with
   ound(cumsum(ms) / frame + 0.5) and derives the word budget from that, so the old per-endpoint truncation was off by one frame at boundaries (e.g. 100 ms notes at 11.609977 ms/frame: 8/9 vs 9/9). It now uses DiffSingerUtils.DurationsMsToFrames over the intervals.
2. Tail SP is its own word. PaddedWordDivAndDur merged the tail SP into the last vowel group (Append(segments.Count - vowelIds[^1])), while the .ds format and DiffSingerScript.ph_num keep it separate (SP sh ir n i SP → 2 2 1 1, not 2 2 2). Runtime pitch/variance therefore had one word fewer than training. The tail segment is now appended as a separate word.

Impact

• dur_type == attn / graphs consuming word_div + word_dur: word_dur now matches the training budget.
• predict_dur == true pitch/variance models: word_div / word_dur now have the same word count as .ds training data.
• �bs duration models and models without word inputs are unaffected.

Verification

• dotnet build OpenUtau.Test: 0 errors.
• DiffSingerWordDivTest (new): 3 passed.
  • [SP, sh, ir, n, i, SP] → wordDiv [2,2,1,1], wordDur [12,9,4,8].
  • no-vowel fallback.
  • cumulative rounding: 100/100 ms at 512/44100 → [9,9].
• Existing OpenUtau.Core.DiffSinger tests: 20 passed, 0 failed.

No changes to the .ds variance clamp (will be a follow-up PR), and this branch is based on upstream master, not on the #2388 revert PR.