Add DuckDB parity checks for point-in-time SQL features

[pavanamthomas]

Addresses #4 by testing the same point-in-time object through two SQL engines and an independent non-SQL calculation.

Implemented:

• deterministic SQLite/DuckDB point-in-time fixture;
• post-cutoff 99999.0 sentinel that correct historical features must exclude;
• independent Pandas calculation of count, spend, and recency;
• cross-engine NULL/no-history checks;
• ROW_NUMBER, RANK, and DENSE_RANK parity on a known tie fixture;
• DuckDB optional dependency exercised in CI;
• docs/cross_engine_parity.md, README, and ROADMAP updates.

Validation: the current PR head passes CI.

Boundaries retained:

• event time is not ingest time;
• PostgreSQL-specific semantics and plans are not tested;
• time zones are not modelled;
• result parity on a deterministic subset is not a claim of full database equivalence.

The point is not another SQL badge. It is whether one information-set definition survives a second SQL engine and an independent Pandas check.