feat(asap-tools): RQE trees and feeder-format data facts in random workloads #881

Part of the RQE frontend tracker. Decisions live in the tracker.

What

In asap-tools/experiments/datasets/fake_metric/random_workloads.py:

1. Every query gets an expr field: its RQE expression tree (shape defined in the A1 issue). Existing query shapes only, no new templates. Leaves are self-contained (id, metric, capability, lookback_ms, grouping, phi/k, value). Queries whose leaves use capabilities outside sum|quantile|cardinality|topk_by_value (the counter-gated topk_by_count) get no expr.
2. data_facts is rewritten in the paced feeder's metrics format (the same format as a sketch-bench eval table's e2e.metrics), so the feeder can generate the data and sketch-bench can plan for it:
   • data: gauge, Pareto α = 2.0, series scaled by Zipf θ = 1.1 over the 10,000 leaves, 10 samples/s per series, 4-level child_of hierarchy (region → service → job → label_0)
   • data_total: counter, same hierarchy/rate/keys
   • events: records table; labels region/service/job/label_0 uniform (skew 0); user_id Zipf θ = 1.1 over 10,000 keys; 10 records/s per leaf series (100,000/s total)
   • the cardinality RQE's metric becomes events (the table its SQL reads)
3. Regenerate and re-commit the seed-862 manifests; keep the byte-identical-for-a-fixed-seed test.
4. Update the design doc (.design_docs/accelerated-random-workload-eval.md) where it describes data_facts.

The leaf JSON must deserialize into A1's type; if A1 has not merged, mirror its field names exactly.

Done when

• generator emits expr + feeder-format facts; tests cover both; manifests regenerated
• the paced feeder accepts the manifest's metrics block (validated by running its parser on one manifest, or a test)