feat: emit dataset statistics report alongside exports (#104)

[dcfocus]

Summary

Follow-up to #96: emits an optional dataset statistics report (<output_path>.stats.json) alongside each export so a training cut is auditable at a glance — without loading the JSONL. Closes #104.

Stacked on #111 (#96) and #112 (#103). Review/merge those first; until then this PR shows their commits too (the stats commit is last).

What it does

• ExportConfig.emit_stats gates a stats artifact computed during the existing export pass (no extra dataset re-materialization).
• ExportStats covers everything the issue asks for:
  • counts: examples, by role, by source, by tenant;
  • token stats (min/median/p95/max/mean) from state_metadata.tokens_used, falling back to a whitespace word-count proxy, with a source field (tokens_used / length_proxy / mixed);
  • grouping: num_groups + records-per-group distribution;
  • curation accounting: records excluded by reason (lifecycle, reward threshold, dedup, decontamination), derived from the curation counts;
  • reward distribution + reward_source breakdown (rollouts), plus the preference form.
• With a train/eval split (Reproducible train/eval split helper for post-training exports #103), each side gets its own .stats.json.

Python

ctx.export_training("training/sft.jsonl", task="sft", emit_stats=True)
# -> training/sft.jsonl.stats.json

Tests

• core (4): role/source/tenant counts + token stats + lifecycle exclusions, length-proxy fallback, rollout reward distribution + reward_source, flag gating.
• python (2): stats contents + flag gating.

README updated.

Checks

• cargo test -p lance-context-core --lib — 66 passed
• cargo fmt --all -- --check, cargo clippy --workspace --all-targets -- -D warnings — clean
• ruff format --check, ruff check, pyright — clean
• test_export_training.py — 12 passed

Acceptance criteria (#104)

• Export optionally emits a stats JSON with the counts/distributions above
• Token stats use state_metadata.tokens_used when available, with a documented fallback (whitespace word count)
• Curation accounting reports excluded counts by reason
• Stats computed in the export pass (no full re-materialization)
• Tests verify counts/distributions against known fixtures

Closes #104