Rewrite avg_over_time as sum over count #410

What to build

Add a semantic-equivalent rewrite for PromQL avg_over_time that exposes mergeable sum and count work, then divides them to preserve the original result.

The existing grouped avg rewrite follows this algebra:

avg(x)  →  float(sum(x)) / count(*)

avg_over_time(x[5m]) is lowered as a Reduction::PerEntity aggregate over the range. It is currently excluded from the rewrite:

let Reduction::Reduce(by) = reduction else {
    return None;
};

After the per-entity multi-measure/schema capability is available, implement the analogous range rewrite while preserving PromQL series identity, labels, timestamps, output name, Float64 type, and null/denominator semantics. The rewrite should make the sum and count portions independently available to existing reuse/realization strategies and return exactly the same observable result as avg_over_time.

Acceptance criteria

• Eligible avg_over_time expressions produce a semantic-equivalent sum/count/divide rewrite candidate.
• The rewritten result preserves input-series identity, labels, timestamps, output name, and Float64 output schema.
• The rewrite is not offered when it could change null or denominator semantics.
• Tests demonstrate equivalence for a representative range query and confirm that the rewrite exposes reusable sum and count work.

Blocked by

• Support multi-measure per-entity aggregate schemas #409 — Per-entity aggregates need multi-measure output-schema support.