Conform forward top-k loss with verl v0.8 - #60

Summary

• add a distinct forward_kl_topk objective and verl-compatible token-mean reduction
• compare values, masses, overlap diagnostics, scalar reduction, and gradients against official verl v0.8.0 at 7aed6b2
• bind cached targets to prompt row, exact actor response tokens, policy version, teacher/tokenizer identity, and score implementation

Validation

• ruff check . and ruff format --check .
• mypy src/miniverl scripts
• actionlint
• 2159 passed, 6 skipped in full CPU suite; 85.76% coverage (one packaging regression found and fixed, focused rerun green)
• 8 GPU tests and 15 network tests passed
• official verl conformance: 2 passed
• strict MkDocs build, package build, and twine checks passed
• frozen calculator JSON remains 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc