Add optional Parquet destinations and Croissant dataset metadata

[ericpaulsen]

Add bounded Parquet output and an MLCommons Croissant sidecar after canonical training-row validation. JSONL and the Sediment manifest remain canonical.

Preserve nested messages, omitted values, source evidence, and recipe semantics. Decision: define how operators supply the dataset's actual license and rights metadata; the software license must not be assumed to license captured data.

Acceptance criteria:

• Physical schemas derive from the canonical schema catalog and preserve supported values losslessly.
• Unsupported unions or types refuse before publication rather than coercing values or inventing sentinels.
• Writers consume bounded batches and have a peak-memory regression.
• Metadata records writer/version, canonical schema identity, row counts, hashes, source bundle identity, and policy.
• Croissant describes resources, fields, splits, declared dataset licensing, and hashes; Sediment Provenance remains in the canonical manifest.
• Round-trip tests preserve semantic rows for supported export objectives; byte equality across writer versions is not promised.
• Optional destination dependencies remain outside the Fact store and Derivation runtime.

Relevant files:

• packages/export/sediment_export/schema_contracts.py
• packages/export/sediment_export/jsonl.py
• packages/export/sediment_export/bounded_training.py
• docs/exports/training-exports.md