[codex] add training regression smoke

[salim4n]

Closes #59

What changed

• Added a focused training regression section to the CI smoke script.
• Exercises ignitionrl train on one discrete environment (GridWorld-v0) and one continuous environment (DroneTarget-v0).
• Checks fixed reward/success thresholds, persisted checkpoint output, and persisted metric series.
• Regression failures include the run id, environment, learner, metric, actual value and expected threshold.

Validation

• bun run smoke:ci
• bun run ci