Make it just work: thinking OFF by default (direct answers), self-verifying launch.sh + chat.sh

- ship chat_template.nothink.jinja; default serves with thinking off so a normal
  request returns content directly (no empty-content reasoning-token trap)
- reasoning parser is conditional: ENABLE_THINKING=1 -> native template + --reasoning-parser glm45
- launch.sh now waits for /health and smoke-tests before printing READY (fails loud otherwise)
- add chat.sh one-liner client; tool calling verified in both modes
- README rewritten around the one-command flow