Fix: report model_is_loaded status immediately instead of waiting up to 10s

[cemalgnlts]

Hi,

This commit reduces the serverless cold start time by ~10 seconds.

Additionally, I've updated the legacy scripts/start_server.sh file to match the one at vast-ai/pyworker/main/start_server.sh.

_model_loaded() sets model_is_loaded=True but doesn't set update_pending,
so _send_metrics_loop can wait up to 10s before notifying report_addr.
This adds update_pending = True to push the status on the next tick (~1s).

Example
Before commit template: cloud.vast.ai/?ref_id=404507&creator_id=404507&name=Qwen3.5 9B - Slow
After commit template: cloud.vast.ai/?ref_id=404507&creator_id=404507&name=Qwen3.5 9B - Fast

Example request template

curl -o /dev/null -s -w "%{time_total} seconds" --request POST \
  --url https://openai.vast.ai/$ENDPOINT/chat/completions \
  --header 'authorization: Bearer $API_KEY' \
  --header 'content-type: application/json' \
  --data '{"messages": [{ "role": "user", "content": "1 + 1 = ?" }]}'

Before commit: 17 - 20 seconds
After commit: 7.9 - 10 seconds