fix(tts): prevent runaway semantic filler loops

[Lucas1479]

Summary

• make static/CUDA-Graph KV attention exclude unused slots and graph-key alignment gaps
• keep CUDA Graph capture sampling-RNG neutral and preserve full prompt/history when a bucket fills
• make FlashAttention2 bucket mode attend only the real contiguous prefix
• add a language/token-ID-independent semantic admission check that retries once before SoVITS/BigVGAN and fails closed if both candidates collapse
• preserve upstream overlay ownership, native infer_panel_naive() generators, streaming inference, and the upstream ref-free idx=0 convention

Closes #3.

Related Amadeus integration: Code-Amadeus/Amadeus#22

Verification

• 92 passed, 1 skipped across tests/
• ruff check aquatts/ tests/
• wheel build succeeded and contains aquatts/inference/semantic_stability.py
• real v3 GPU matrix: 4 filler/control texts x 20 paired seeds x Graph/static/dynamic = 240 trials
  • before: 27 collapsed rows; 78/80 text-seed groups differed across modes
  • after: 0 runtime errors, 0 final collapses, 0/80 token-hash mismatches
  • 10 fault seeds were rejected and recovered once in all three modes
  • median semantic time: Graph 0.10 s, static 0.33 s, dynamic 0.33 s
• end-to-end seed 3003: rejected 600-token candidate (equal/periodic run 186), recovered to 113 tokens, rendered 5.121 s audio

Safety / compatibility

• no vendored t2s_model.py is restored; the fix targets the latest in-memory upstream overlay
• no hard-coded language or semantic token IDs
• guard can be disabled with AQUATTS_SEMANTIC_GUARD=0 for diagnosis
• invalid candidates are never cached or passed to the vocoder