Add first-class Parquet prompt OPD runtime - #59

Summary

• add a discriminated environment / verl_parquet RunConfig source without silently substituting a tool environment
• stream and validate verl-style Parquet prompts with row, schema, content, tokenizer, and rendered-prompt provenance
• add source-agnostic rollout runtimes and real masked padded greedy prompt generation
• train pure OPD end to end without an environment or reward; response tokens are the only selected labels
• preserve the existing multi-turn ToolEnvironment path and fail closed on unsupported reward/oracle/privileged-context semantics

Validation

• 2154 passed, 6 skipped, 21 deselected; 85.76% coverage
• GPU: 8 passed on RTX 4080
• network: 15 passed
• Ruff check/format, mypy, actionlint
• mkdocs build --strict
• package build and twine check
• calculator frozen JSON SHA-256 remains 53fc1d4d5b7adee09618d77ad62d4086ba56b78569832d6fc7c3bcd5c2695bbc

Scope

This is PR B of the v0.8.0 sequence. It does not change frozen scientific results or claim task reward support for prompt datasets.