Support multi-measure per-entity aggregate schemas

[milindsrivastava1997]

What to build

Allow a Reduction::PerEntity aggregate to represent more than one aggregate measure while preserving the label set, timestamp, and a well-defined output schema for every measure.

Today, per-entity range reductions are structurally limited to one measure. This blocks algebraically valid rewrites such as avg_over_time(x) → sum_over_time(x) / count_over_time(x), tracked by #410.

Current schema derivation encodes the limitation:

Reduction::PerEntity => {
    debug_assert_eq!(measures.len(), 1, "a per-entity reduction is single-aggregate");
    return Ok(per_series_reduction_schema(in_schema, &measures[0]));
}

and the per-series path replaces the one conventional sample-value column:

let mut columns = input.columns.clone();
columns[vi] = out; // `out.name` is forced to "value"

The resulting schema cannot carry both a per-series sum and count. Define and implement the smallest compatible multi-measure schema contract, including naming and type behavior, so downstream lowering and binary operations can safely consume it.

Acceptance criteria

• A PerEntity aggregate with multiple measures has a deterministic, validated output schema that preserves timestamp and labels and exposes every measure.
• The single-measure PromQL range-function schema remains unchanged: its output sample column is named value and is Float64.
• Schema derivation, resolution, and physical lowering either support the new shape end-to-end or reject unsupported forms with a clear error; they must not silently drop measures.
• Focused tests cover a two-measure per-entity aggregate and protect existing single-measure range-function behavior.

Blocked by

None — can start immediately.