Promote shared MOSS runtime pieces

[0xShug0]

Summary

• promote the MOSS audio tokenizer codec runtime into framework code so MOSS-family models share one lifecycle-managed runtime
• replace duplicated MOSS sampling code with the shared HF sampler
• move the MOSS token-row builder into the codec runtime helper surface and remove the old model-local shared files
• clean stale MOSS VoiceGen test includes/comments after the shared helper move

Validation

• conda run --no-capture-output -n qwen3-tts cmake --build build/debug --parallel $(nproc) --target audiocpp_cli
• conda run --no-capture-output -n qwen3-tts cmake --build build/debug --parallel $(nproc) --target moss_tts_local_smoke moss_tts_local_warm_bench moss_tts_nano_warm_bench moss_voicegen_smoke moss_voicegen_prompt_parity moss_voicegen_backbone_parity moss_voicegen_generation_parity moss_voicegen_codec_parity
• path rerun vs pre-token-row-fold baseline B:
  • moss_tts_local_long_lived_session: artifacts/text match, +2.73%
  • moss_tts_nano_sampled_multi_request: artifacts/text match, +1.72%
  • moss_voicegen_radio_voice_design: artifacts/text match, +3.00%
• current branch vs main HEAD:
  • moss_tts_local: artifacts/text match
  • moss_tts_nano: artifacts/text match
  • vietneu_tts: byte-identical WAV, SHA256 812ae328819fc14fd6e43e21980c9dd61e273727c93e84b58015e2218f19c369
  • moss_voicegen: changed output, expected/allowed for this refactor path