Add pd disaggregated inference

[Bihan]

Testing Steps

1. Create (CPU node) in K8s cluster

2. Create gateway in the CPU node using below config

type: gateway
name: bihan-gateway

backend: kubernetes
region: any

domain: bihan-gateway.dstack.ai
router:
  type: sglang

3. Create GPU-node with 3 instances (1 Prefill, 1 Decode and 1 for testing scaling) in the same K8s cluster where gateway node exists.
   Note: See design doc for details on why the gateway and workers are required to be on the same network.

4. Apply below prefill-decode service configuration

type: service
name: prefill-decode
image: lmsysorg/sglang:latest

env:
  - HF_TOKEN
  - MODEL_ID=meta-llama/Llama-3.2-1B-Instruct

replicas:
  - count: 1..2
    scaling:
      metric: rps
      target: 3
    commands:
      - |
          python -m sglang.launch_server \
            --model-path $MODEL_ID \
            --disaggregation-mode prefill \
            --disaggregation-transfer-backend mooncake \
            --host 0.0.0.0 \
            --port 8000 \
            --disaggregation-bootstrap-port 8998 \
            --log-level debug \
            > worker-server.log 2>&1
    resources:
      gpu: 1

  - count: 1
    commands:
      - |
          python -m sglang.launch_server \
            --model-path $MODEL_ID \
            --disaggregation-mode decode \
            --disaggregation-transfer-backend mooncake \
            --host 0.0.0.0 \
            --port 8000 \
            --log-level debug \
            > worker-server.log 2>&1
    resources:
      gpu: 1

port: 8000
model: meta-llama/Llama-3.2-1B-Instruct

probes:
  - type: http
    url: /health_generate
    interval: 15s

router_config:
  policy: round_robin
  pd_disaggregation: true

5. When rps>=3 prefill replica scales to 2.

Note: For testing you need to assign wheel to https://bihan-test-bucket.s3.eu-west-1.amazonaws.com/dstack_gateway-0.0.1-py3-none-any.whl