Perf(qwen3-tts): fuse CodePredictor RoPE and reuse decode masks

[qaqjx]

What does this PR do?

Optimizes the Qwen3-TTS autoregressive generation path.

• Replaces the repeated PyTorch RoPE operations in CodePredictor with the fused Triton apply_rope_pos_ids kernel.
• Reuses the static invalid-token mask during decode and passes only the per-request suppress_eos flag into the CUDA Graph.
• Adds correctness coverage and a deterministic fixed-frame benchmark.

On an NVIDIA H20, CodePredictor CUDA Graph latency decreased by 35.8-36.9%, end-to-end latency decreased by 4.1-4.4%, and worker GPU memory decreased by about 1.5%.

How was it tested?

• .venv/bin/ruff check .
• .venv/bin/pytest -q test/modular/test_qwen3_tts_model.py (35 passed)
• CUDA_VISIBLE_DEVICES=0 HF_HUB_OFFLINE=1 TRANSFORMERS_OFFLINE=1 .venv/bin/pytest -q test/integration/test_qwen3_tts_real_weights.py (5 passed)
• End-to-end fixed-frame benchmarks for 32, 64, and 128 frames, plus concurrency 2, 4, and 8 and natural-EOS validation.

Performance results

Measured with Qwen/Qwen3-TTS-12Hz-0.6B-CustomVoice on one NVIDIA H20:

Metric	Baseline	Optimized	Change
CodePredictor CUDA Graph, BS=1	10.077 ms	6.359 ms	-36.9%
End-to-end latency, 32 frames	333.1 ms	318.6 ms	-4.3%
End-to-end latency, 64 frames	587.4 ms	561.7 ms	-4.4%
End-to-end latency, 128 frames	1,146.4 ms	1,099.7 ms	-4.1%

The CodePredictor improvement was 35.8-36.9% across BS=1/2/4/8. At concurrency 2, throughput improved by 2.9%; concurrency 4 and 8 were saturated and showed no statistically clear gain or regression. Three natural-EOS requests produced non-empty audio and stopped before talker_max_tokens.

Checklist

• ruff check . passes
• Added or updated tests / docs where relevant

[NSagan271]

Overall looks good, just one comment. Also @qaqjx #201 (the multi-sampler PR) got merged, so can you rebase off of main and re-request my review? That PR already migrated Qwen TTS to the new sampler configuration.

[NSagan271]

LGTM!