fix(server): make uvicorn workers>1 work (factory + spawn for CUDA)

[vladlearns]

FIXES:
#1142
#1143

Bugs sum up

1. On linux, default multiprocessing method is fork(). CUDA ctxs can't be inherited by forked child - when multiple workers try to init CUDA after forking, they crash. Only one worker survives /w uv retries;
2. Got an issues running /w workers > 1 in 1 container, each worker starts as a separate process and imports the module fresh - cli args parsed in __main__ only exist in the parent process, so workers would miss that config and fall back to defaults/break.

What I changed

• Set multiprocessing to spawn before starting uv, so each worker starts fresh = no fork = no CUDA init crashes;
• Parent now dumps parsed args to json and stores them in var before starting uv;
• Workers calls the factory, it reads the var, loads the json, and rebuilds the namespace, so every worker gets the same config.

QA

• Solo-worker - same;
• Multi-workers - config is passed to all workers;
• Tested in Docker with 3 workers on 4090.

[vladlearns]

@Juanma-t, this should help you

[github-actions]

This PR is stale because it has been open for 30 days with no activity.