Qwen3-TTS PR 4: rank_devices, several workers on one GPU

[merceod]

Adds an optional rank_devices: {rank: device_index} key to deployment configs. A worker is one process per rank and rank n runs on device n by default. The mapping lets two node groups run as separate workers on the same GPU:

node_groups:
  - {node_names: [Talker], ranks: [0]}
  - {node_names: [Codec], ranks: [1]}
rank_devices: {1: 0}

Why: a single worker loop alternates the steps of every node it hosts, so for a streaming TTS graph every codec window sits between two Talker steps.

What it does for Qwen3-TTS 1.7B on one H100 (numbers and the config in the model PR). On its own the split lowers first-audio p50 by 15-30% at c>=8 with throughput within a few percent of one worker, because without CUDA MPS two processes time-slice the GPU. Under a user-level MPS daemon (nvidia-cuda-mps-control -d before mstar serve) the workers' kernels overlap and c=32 goes from 153 ms / 76 audio-s/s to 75 ms / 91 audio-s/s (70 ms / 100 audio-s/s with a bf16 codec). The split config caps the Talker KV pool at 1024 pages so both workers fit in memory.

Three commits, one per file (conductor.py, docs/serving.rst, test/modular/test_worker_device_map.py). Deployments without the key are unchanged. Full test/modular passes.