[Serve] Enable multiple ports in SkyServe replicas

[Conless]

Current implementation of SkyServe only allows the replicas to expose one port. In some cases, services may need to expose multiple ports for custom controller, GUI interface, etc. This PR adds support for it by allowing multiple ports, but requiring user to specify one as the main port in the service section, if multiple ports are provided.

service:
  readiness_probe: /
  replicas: 2
  port: 8000

resources:
  ports:
  - 8000-8001
  - 3000
  cpus: 2

The output of sky serve status will look like:

Services
NAME              VERSION  UPTIME  STATUS  REPLICAS  ENDPOINT            
sky-service-xxxx  1        28s     READY   2/2       xxxx:30001  

Service Replicas
SERVICE_NAME      ID  VERSION  ENDPOINT             LAUNCHED     RESOURCES       STATUS  REGION     
sky-service-xxxx  1   1        http://xxxx:8000     50 secs ago  1x AWS(vCPU=2)  READY   us-east-1  
sky-service-xxxx  2   1        http://xxxx:8000     1 min ago    1x AWS(vCPU=2)  READY   us-east-1 

while the other ports (8001, 3000) are still accessible.

Tested (run the relevant ones):

• Code formatting: bash format.sh
• Any manual or new tests for this PR (please specify below)
  • sky serve up, sky serve status and accessing the service with config:

  service:
    readiness_probe: /
    replicas: 2
    port: 8000
  
  resources:
    ports:
    - 8000-8001
    - 3000
    cpus: 2
  
  run: |
    python -m http.server 8000 &
    python -m http.server 8001 &
    python -m http.server 3000 &

• All smoke tests: pytest tests/test_smoke.py
• Relevant individual smoke tests: pytest tests/test_smoke.py::test_fill_in_the_name
• Backward compatibility tests: conda deactivate; bash -i tests/backward_compatibility_tests.sh

[Conless]

Hi @cblmemo ! Would you like to have a look at this?

[cblmemo]

Thanks for adding this feature @Conless ! Mostly LGTM. Left a discussion.

Also we might want to update related documents for this :))

[cblmemo]

Can we also update the PR description?

[cblmemo]

Also, it will be great if we can test on some real world usage (e.g. deploy an LLM service and expose metric port in vLLM or dashboard in ray