`feat` Expression group keys — temporal bucketing (trunc) for grouped aggregates

[0x054]

What to build

An expression-capable group key for aggregate projections, temporal first: a portable truncation surface — t.date.trunc("month") (or ferro.trunc("month", t.date)) — compiling to date_trunc on Postgres, usable as a projected field, group key, and order_by source exactly like a plain column.

Requested by Pinch (M8 reporting, PRD 0012, found at the M8 CP0 spike pinch-backend#46). Blocks nothing — recorded demand, not a gate: M8 ships on daily-grain GROUP BY t.date (verified working on 0.17.1) + Python folding, which is fine while fetched rows = days-with-activity. This retires that fold when key spaces grow, and is the first concrete consumer of the record plan's reserved per-field expr slot (ADR-0007/0009).

The gap, verified on 0.17.1

Group keys are columns only — plain or traversed. No date-part accessor or function surface exists in the query API (src/ferro/query/, aggregation guide). Wanted:

rows = await Transaction.select(
    lambda t: {
        "month": t.date.trunc("month"),   # ← does not exist
        "total": t.amount_minor.sum(),
    }
).where(lambda t: t.amount_minor < 0).all()

Acceptance criteria

• trunc granularities day/week/month/quarter/year on date/datetime columns, both as projected field and group key; order_by composes.
• Build-time rejection of granularities/types with no portable meaning (the guide's build-time-honesty stance).
• Record-plan expr field shape exercised end to end; golden vectors extended.

Out of scope

General computed expressions, HAVING, window functions — temporal truncation is the whole ask.