docs(moss-tts): document MOSS-TTS-Local variant in cookbook

[xinlij]

Summary

Adds the MOSS-TTS-Local-Transformer-v1.5 variant to the MOSS-TTS cookbook (docs/cookbook/moss_tts.md), which previously only documented the delay model (MOSS-TTS-v1.5).

Prerequisites

• Distinguishes the two checkpoints (single-GPU delay model vs. two-GPU Local variant).
• Adds the hf download OpenMOSS-Team/MOSS-TTS-Local-Transformer-v1.5 command.
• Notes the codec (MOSS-Audio-Tokenizer-v2, 48 kHz output) is loaded via the checkpoint's remote code, so trust_remote_code must be enabled.

Server Configuration

• Splits into two subsections, one per checkpoint.
• Documents the Local launch with examples/configs/moss_tts_local.yaml and its two-GPU default (AR engine on cuda:0, codec on cuda:1).
• Notes the single-GPU option via config_cls: MossTTSLocalColocatedPipelineConfig.

The request shape and generation parameters are identical to the delay model, so the existing synthesis examples apply to both variants. Note the output sample rate differs (Local 48 kHz vs. delay 24 kHz).

Verified against the codebase

• Architecture (36-layer Qwen3 backbone + 1-layer frame-local transformer): sglang_omni/models/moss_tts_local/sglang_model.py.
• Two-GPU default (codec_device="cuda:1") vs. colocated (codec_device="cuda:0"): sglang_omni/models/moss_tts_local/config.py.
• config_cls: MossTTSLocalColocatedPipelineConfig resolves via the Variants lookup in sglang_omni/models/registry.py (get_config_cls_by_name).
• 48 kHz output: sglang_omni/models/moss_tts_local/payload_types.py (sample_rate = 48000).

[gaoyang07]

For readability and discoverability, we might consider splitting this section into two dedicated subsections from the beginning: one for the Delay Pattern model and one for the Local Transformer model. Each subsection could describe the corresponding architecture, serving pipeline, hardware assumptions, and usage examples. This would make the tradeoff between the two token modeling patterns more explicit and avoid mixing model-specific details in a single flow.