🐛 Make RL circuit sampling reproducible

[flowerthrower]

🤖 AI text below 🤖

Description

Makes RL training-circuit sampling honor Gymnasium's seed lifecycle.

PredictorEnv now uses Gymnasium's np_random, so reset(seed=...) controls sampling, and candidate QASM paths are sorted before indexed selection to make identical seeds reproducible across filesystems.

This draft is stacked on the TKET pass PR. It does not change BQSKit's established fixed pass seed or unrelated ML estimators. Validated with repository lint and focused sampling smoke checks.

Part of #664

Checklist

• The diff is focused and reuses Gymnasium's RNG.
• Lint passes.
• I confirm that I have personally reviewed and understood all AI-generated content.

If PR contains AI-assisted content:

• AI assistance was authorized and is disclosed.
• This body begins with the required visible disclosure.

[flowerthrower]

🤖 AI text below 🤖\n\nSuperseded while rebuilding the corrected downstream stack after #758.