Update docs to help agents handle full lifecycle

[myurasov-nv]

Docs-only: three gaps that caused real failures while an AI agent drove full GPU-run lifecycles (create -> provision -> train -> collect -> delete). Verified 2026-07-25 on launchpad 2x H100, verda B300, massedcompute L40S, GCP T4.

Issues

Problem observed	Change
brev exec "nohup train.py &" (ml-training.md step 7) blocks for the whole run - exec holds the session even for nohup children; step 8 tail -f never exits	Detached-launch pattern (setsid nohup ... & disown via @launch.sh + completion marker to poll); tail -n instead of tail -f
Readiness lifecycle undocumented: RUNNING != SHELL READY; sshd refuses 1-2 min after READY; benign ErrorForbidden from the create poll	Lifecycle note under Create Instances
Billing model absent: per-second, RUNNING-only (BUILDING free, no round-up); stopped instances keep accruing storage; usage feed lags	"How Billing Works" section in SKILL.md + warning in cleanup.md

Files

• SKILL.md: detach pattern, lifecycle note, billing section
• reference/commands.md: long-running note under brev exec
• prompts/ml-training.md: fix background launch, monitoring
• prompts/cleanup.md: stopped-instances-cost warning