Let llama.cpp own boot warmup

[SoryAK]

Summary

• Stop MBA from POSTing /completion after llama.cpp is healthy (that 60s generate timed out on large models and killed the server).
• Emit llama.cpp --warmup / --no-warmup from warmupTokens so the empty run happens during load.
• Boot returns when /health is ok.

Test plan

• systemctl --user restart mba so the daemon picks up the source change
• mba s boot a model that previously failed with warmup failed: TimeoutError
• CLI prints (waits for health)… and the spawn argv includes --warmup
• Boot completes without MBA sending /completion

Made with Cursor