[Bug]: [HiggsAudioV3] Talker crashes CUDA graph capture (low_latency profile) due to boolean-mask indexing

[jajmangold]

Describe the bug

The Higgs-Audio-v3 Stage-0 talker crashes during CUDA graph capture when the
higgs_multimodal_qwen3_low_latency profile is used (Stage-0
enforce_eager: false, cudagraph_mode: FULL_DECODE_ONLY).

HiggsAudioV3Talker._decode_request_token_positions filters its result with
boolean-mask indexing:

decode_mask = (spans == 1) & (starts >= 0) & (starts < int(num_tokens))
return req_rows[decode_mask], starts[decode_mask]

req_rows[decode_mask] produces a data-dependent output shape, which forces
a host synchronization and is illegal during CUDA graph stream capture. This is
backend-independent (it is a CUDA stream-capture rule, not specific to any
attention backend).

Repro

1. Deploy higgs-audio-v3-tts-4b with the Stage-0 CUDA graph profile
   (deploy/higgs_multimodal_qwen3_low_latency.yaml:
   enforce_eager: false, compilation_config.cudagraph_mode: FULL_DECODE_ONLY).
2. Run any TTS request.
3. Capture fails:

torch.AcceleratorError: CUDA error: operation not permitted when stream is capturing
  File ".../higgs_audio_v3_talker.py", in forward
    hidden_states = self._apply_audio_feedback(hidden_states, input_ids)
  File ".../higgs_audio_v3_talker.py", in _apply_audio_feedback
    req_rows, token_positions = self._decode_request_token_positions(num_tokens, hidden_states.device)
  File ".../higgs_audio_v3_talker.py", in _decode_request_token_positions
    return req_rows[decode_mask], starts[decode_mask]

Observed on Tesla V100 / sm70 (FLASH_ATTN_V100 backend), but the boolean-mask is
backend-independent.

Fix

Under the external decode CUDA graph the decode batch is always a uniform
single-token decode (every span == 1), so decode_mask is all-True and the
filtered result equals the unfiltered tensors. Returning them directly when
_use_external_decode_cudagraph is set keeps the captured shape static and fixes
the crash, with no change to the eager path. A PR follows.